Migrated from GitHub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+docs/teaser.gif filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,352 @@
+Creative Commons Attribution-NonCommercial 4.0 International
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright and
+certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+-   Considerations for licensors: Our public licenses are intended for
+    use by those authorized to give the public permission to use
+    material in ways otherwise restricted by copyright and certain other
+    rights. Our licenses are irrevocable. Licensors should read and
+    understand the terms and conditions of the license they choose
+    before applying it. Licensors should also secure all rights
+    necessary before applying our licenses so that the public can reuse
+    the material as expected. Licensors should clearly mark any material
+    not subject to the license. This includes other CC-licensed
+    material, or material used under an exception or limitation to
+    copyright. More considerations for licensors :
+    wiki.creativecommons.org/Considerations_for_licensors
+
+-   Considerations for the public: By using one of our public licenses,
+    a licensor grants the public permission to use the licensed material
+    under specified terms and conditions. If the licensor's permission
+    is not necessary for any reason–for example, because of any
+    applicable exception or limitation to copyright–then that use is not
+    regulated by the license. Our licenses grant only permissions under
+    copyright and certain other rights that a licensor has authority to
+    grant. Use of the licensed material may still be restricted for
+    other reasons, including because others have copyright or other
+    rights in the material. A licensor may make special requests, such
+    as asking that all changes be marked or described. Although not
+    required by our licenses, you are encouraged to respect those
+    requests where reasonable. More considerations for the public :
+    wiki.creativecommons.org/Considerations_for_licensees
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and conditions.
+
+-   Section 1 – Definitions.
+
+    -   a. Adapted Material means material subject to Copyright and
+        Similar Rights that is derived from or based upon the Licensed
+        Material and in which the Licensed Material is translated,
+        altered, arranged, transformed, or otherwise modified in a
+        manner requiring permission under the Copyright and Similar
+        Rights held by the Licensor. For purposes of this Public
+        License, where the Licensed Material is a musical work,
+        performance, or sound recording, Adapted Material is always
+        produced where the Licensed Material is synched in timed
+        relation with a moving image.
+    -   b. Adapter's License means the license You apply to Your
+        Copyright and Similar Rights in Your contributions to Adapted
+        Material in accordance with the terms and conditions of this
+        Public License.
+    -   c. Copyright and Similar Rights means copyright and/or similar
+        rights closely related to copyright including, without
+        limitation, performance, broadcast, sound recording, and Sui
+        Generis Database Rights, without regard to how the rights are
+        labeled or categorized. For purposes of this Public License, the
+        rights specified in Section 2(b)(1)-(2) are not Copyright and
+        Similar Rights.
+    -   d. Effective Technological Measures means those measures that,
+        in the absence of proper authority, may not be circumvented
+        under laws fulfilling obligations under Article 11 of the WIPO
+        Copyright Treaty adopted on December 20, 1996, and/or similar
+        international agreements.
+    -   e. Exceptions and Limitations means fair use, fair dealing,
+        and/or any other exception or limitation to Copyright and
+        Similar Rights that applies to Your use of the Licensed
+        Material.
+    -   f. Licensed Material means the artistic or literary work,
+        database, or other material to which the Licensor applied this
+        Public License.
+    -   g. Licensed Rights means the rights granted to You subject to
+        the terms and conditions of this Public License, which are
+        limited to all Copyright and Similar Rights that apply to Your
+        use of the Licensed Material and that the Licensor has authority
+        to license.
+    -   h. Licensor means the individual(s) or entity(ies) granting
+        rights under this Public License.
+    -   i. NonCommercial means not primarily intended for or directed
+        towards commercial advantage or monetary compensation. For
+        purposes of this Public License, the exchange of the Licensed
+        Material for other material subject to Copyright and Similar
+        Rights by digital file-sharing or similar means is NonCommercial
+        provided there is no payment of monetary compensation in
+        connection with the exchange.
+    -   j. Share means to provide material to the public by any means or
+        process that requires permission under the Licensed Rights, such
+        as reproduction, public display, public performance,
+        distribution, dissemination, communication, or importation, and
+        to make material available to the public including in ways that
+        members of the public may access the material from a place and
+        at a time individually chosen by them.
+    -   k. Sui Generis Database Rights means rights other than copyright
+        resulting from Directive 96/9/EC of the European Parliament and
+        of the Council of 11 March 1996 on the legal protection of
+        databases, as amended and/or succeeded, as well as other
+        essentially equivalent rights anywhere in the world.
+    -   l. You means the individual or entity exercising the Licensed
+        Rights under this Public License. Your has a corresponding
+        meaning.
+
+-   Section 2 – Scope.
+
+    -   a. License grant.
+        -   1. Subject to the terms and conditions of this Public
+            License, the Licensor hereby grants You a worldwide,
+            royalty-free, non-sublicensable, non-exclusive, irrevocable
+            license to exercise the Licensed Rights in the Licensed
+            Material to:
+            -   A. reproduce and Share the Licensed Material, in whole
+                or in part, for NonCommercial purposes only; and
+            -   B. produce, reproduce, and Share Adapted Material for
+                NonCommercial purposes only.
+        -   2. Exceptions and Limitations. For the avoidance of doubt,
+            where Exceptions and Limitations apply to Your use, this
+            Public License does not apply, and You do not need to comply
+            with its terms and conditions.
+        -   3. Term. The term of this Public License is specified in
+            Section 6(a).
+        -   4. Media and formats; technical modifications allowed. The
+            Licensor authorizes You to exercise the Licensed Rights in
+            all media and formats whether now known or hereafter
+            created, and to make technical modifications necessary to do
+            so. The Licensor waives and/or agrees not to assert any
+            right or authority to forbid You from making technical
+            modifications necessary to exercise the Licensed Rights,
+            including technical modifications necessary to circumvent
+            Effective Technological Measures. For purposes of this
+            Public License, simply making modifications authorized by
+            this Section 2(a)(4) never produces Adapted Material.
+        -   5. Downstream recipients.
+            -   A. Offer from the Licensor – Licensed Material. Every
+                recipient of the Licensed Material automatically
+                receives an offer from the Licensor to exercise the
+                Licensed Rights under the terms and conditions of this
+                Public License.
+            -   B. No downstream restrictions. You may not offer or
+                impose any additional or different terms or conditions
+                on, or apply any Effective Technological Measures to,
+                the Licensed Material if doing so restricts exercise of
+                the Licensed Rights by any recipient of the Licensed
+                Material.
+        -   6. No endorsement. Nothing in this Public License
+            constitutes or may be construed as permission to assert or
+            imply that You are, or that Your use of the Licensed
+            Material is, connected with, or sponsored, endorsed, or
+            granted official status by, the Licensor or others
+            designated to receive attribution as provided in Section
+            3(a)(1)(A)(i).
+    -   b. Other rights.
+        -   1. Moral rights, such as the right of integrity, are not
+            licensed under this Public License, nor are publicity,
+            privacy, and/or other similar personality rights; however,
+            to the extent possible, the Licensor waives and/or agrees
+            not to assert any such rights held by the Licensor to the
+            limited extent necessary to allow You to exercise the
+            Licensed Rights, but not otherwise.
+        -   2. Patent and trademark rights are not licensed under this
+            Public License.
+        -   3. To the extent possible, the Licensor waives any right to
+            collect royalties from You for the exercise of the Licensed
+            Rights, whether directly or through a collecting society
+            under any voluntary or waivable statutory or compulsory
+            licensing scheme. In all other cases the Licensor expressly
+            reserves any right to collect such royalties, including when
+            the Licensed Material is used other than for NonCommercial
+            purposes.
+
+-   Section 3 – License Conditions.
+
+    Your exercise of the Licensed Rights is expressly made subject to
+    the following conditions.
+
+    -   a. Attribution.
+        -   1. If You Share the Licensed Material (including in modified
+            form), You must:
+            -   A. retain the following if it is supplied by the
+                Licensor with the Licensed Material:
+                -   i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if designated);
+                -   ii. a copyright notice;
+                -   iii. a notice that refers to this Public License;
+                -   iv. a notice that refers to the disclaimer of
+                    warranties;
+                -   v. a URI or hyperlink to the Licensed Material to
+                    the extent reasonably practicable;
+            -   B. indicate if You modified the Licensed Material and
+                retain an indication of any previous modifications; and
+            -   C. indicate the Licensed Material is licensed under this
+                Public License, and include the text of, or the URI or
+                hyperlink to, this Public License.
+        -   2. You may satisfy the conditions in Section 3(a)(1) in any
+            reasonable manner based on the medium, means, and context in
+            which You Share the Licensed Material. For example, it may
+            be reasonable to satisfy the conditions by providing a URI
+            or hyperlink to a resource that includes the required
+            information.
+        -   3. If requested by the Licensor, You must remove any of the
+            information required by Section 3(a)(1)(A) to the extent
+            reasonably practicable.
+        -   4. If You Share Adapted Material You produce, the Adapter's
+            License You apply must not prevent recipients of the Adapted
+            Material from complying with this Public License.
+
+-   Section 4 – Sui Generis Database Rights.
+
+    Where the Licensed Rights include Sui Generis Database Rights that
+    apply to Your use of the Licensed Material:
+
+    -   a. for the avoidance of doubt, Section 2(a)(1) grants You the
+        right to extract, reuse, reproduce, and Share all or a
+        substantial portion of the contents of the database for
+        NonCommercial purposes only;
+    -   b. if You include all or a substantial portion of the database
+        contents in a database in which You have Sui Generis Database
+        Rights, then the database in which You have Sui Generis Database
+        Rights (but not its individual contents) is Adapted Material;
+        and
+    -   c. You must comply with the conditions in Section 3(a) if You
+        Share all or a substantial portion of the contents of the
+        database.
+
+    For the avoidance of doubt, this Section 4 supplements and does not
+    replace Your obligations under this Public License where the
+    Licensed Rights include other Copyright and Similar Rights.
+
+-   Section 5 – Disclaimer of Warranties and Limitation of Liability.
+
+    -   a. Unless otherwise separately undertaken by the Licensor, to
+        the extent possible, the Licensor offers the Licensed Material
+        as-is and as-available, and makes no representations or
+        warranties of any kind concerning the Licensed Material, whether
+        express, implied, statutory, or other. This includes, without
+        limitation, warranties of title, merchantability, fitness for a
+        particular purpose, non-infringement, absence of latent or other
+        defects, accuracy, or the presence or absence of errors, whether
+        or not known or discoverable. Where disclaimers of warranties
+        are not allowed in full or in part, this disclaimer may not
+        apply to You.
+    -   b. To the extent possible, in no event will the Licensor be
+        liable to You on any legal theory (including, without
+        limitation, negligence) or otherwise for any direct, special,
+        indirect, incidental, consequential, punitive, exemplary, or
+        other losses, costs, expenses, or damages arising out of this
+        Public License or use of the Licensed Material, even if the
+        Licensor has been advised of the possibility of such losses,
+        costs, expenses, or damages. Where a limitation of liability is
+        not allowed in full or in part, this limitation may not apply to
+        You.
+    -   c. The disclaimer of warranties and limitation of liability
+        provided above shall be interpreted in a manner that, to the
+        extent possible, most closely approximates an absolute
+        disclaimer and waiver of all liability.
+
+-   Section 6 – Term and Termination.
+
+    -   a. This Public License applies for the term of the Copyright and
+        Similar Rights licensed here. However, if You fail to comply
+        with this Public License, then Your rights under this Public
+        License terminate automatically.
+    -   b. Where Your right to use the Licensed Material has terminated
+        under Section 6(a), it reinstates:
+
+        -   1. automatically as of the date the violation is cured,
+            provided it is cured within 30 days of Your discovery of the
+            violation; or
+        -   2. upon express reinstatement by the Licensor.
+
+        For the avoidance of doubt, this Section 6(b) does not affect
+        any right the Licensor may have to seek remedies for Your
+        violations of this Public License.
+
+    -   c. For the avoidance of doubt, the Licensor may also offer the
+        Licensed Material under separate terms or conditions or stop
+        distributing the Licensed Material at any time; however, doing
+        so will not terminate this Public License.
+    -   d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+        License.
+
+-   Section 7 – Other Terms and Conditions.
+
+    -   a. The Licensor shall not be bound by any additional or
+        different terms or conditions communicated by You unless
+        expressly agreed.
+    -   b. Any arrangements, understandings, or agreements regarding the
+        Licensed Material not stated herein are separate from and
+        independent of the terms and conditions of this Public License.
+
+-   Section 8 – Interpretation.
+
+    -   a. For the avoidance of doubt, this Public License does not, and
+        shall not be interpreted to, reduce, limit, restrict, or impose
+        conditions on any use of the Licensed Material that could
+        lawfully be made without permission under this Public License.
+    -   b. To the extent possible, if any provision of this Public
+        License is deemed unenforceable, it shall be automatically
+        reformed to the minimum extent necessary to make it enforceable.
+        If the provision cannot be reformed, it shall be severed from
+        this Public License without affecting the enforceability of the
+        remaining terms and conditions.
+    -   c. No term or condition of this Public License will be waived
+        and no failure to comply consented to unless expressly agreed to
+        by the Licensor.
+    -   d. Nothing in this Public License constitutes or may be
+        interpreted as a limitation upon, or waiver of, any privileges
+        and immunities that apply to the Licensor or You, including from
+        the legal processes of any jurisdiction or authority.
+
+Creative Commons is not a party to its public licenses. Notwithstanding,
+Creative Commons may elect to apply one of its public licenses to
+material it publishes and in those instances will be considered the
+"Licensor." The text of the Creative Commons public licenses is
+dedicated to the public domain under the CC0 Public Domain Dedication.
+Except for the limited purpose of indicating that material is shared
+under a Creative Commons public license or as otherwise permitted by the
+Creative Commons policies published at creativecommons.org/policies,
+Creative Commons does not authorize the use of the trademark "Creative
+Commons" or any other trademark or logo of Creative Commons without its
+prior written consent including, without limitation, in connection with
+any unauthorized modifications to any of its public licenses or any
+other arrangements, understandings, or agreements concerning use of
+licensed material. For the avoidance of doubt, this paragraph does not
+form part of the public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

ORIGINAL_README.md

@@ -0,0 +1,140 @@
+# MangaNinja: Line Art Colorization with Precise Reference Following
+
+This repository represents the official implementation of the paper titled "MangaNinja: Line Art Colorization with Precise Reference Following".
+
+[![Website](docs/badge-website.svg)](https://johanan528.github.io/MangaNinjia/)
+[![Paper](https://img.shields.io/badge/arXiv-PDF-b31b1b)](https://arxiv.org/abs/2501.08332)
+[![License](https://img.shields.io/badge/License-CC%20BY--NC%204.0-929292)](https://creativecommons.org/licenses/by-nc/4.0/)
+
+<p align="center">
+    <a href="https://johanan528.github.io/"><strong>Zhiheng Liu*</strong></a>
+    ·
+    <a href="https://felixcheng97.github.io/"><strong>Ka Leong Cheng*</strong></a>
+    ·
+    <a href="https://xavierchen34.github.io/"><strong>Xi Chen</strong></a>
+    ·
+    <a href="https://jiexiaou.github.io/"><strong>Jie Xiao</strong></a>
+    ·
+    <a href="https://ken-ouyang.github.io/"><strong>Hao Ouyang</strong></a>
+    ·
+    <a href="https://scholar.google.com/citations?user=Mo_2YsgAAAAJ&hl=zh-CN"><strong>Kai Zhu</strong></a>
+    ·
+    <a href="https://scholar.google.com/citations?user=8zksQb4AAAAJ&hl=zh-CN"><strong>Yu Liu</strong></a>
+    ·
+    <a href="https://shenyujun.github.io/"><strong>Yujun Shen</strong></a>
+    ·
+    <a href="https://cqf.io/"><strong>Qifeng Chen</strong></a>
+    ·
+    <a href="http://luoping.me/"><strong>Ping Luo</strong></a>
+    <br>
+  </p>
+
+We propose **MangaNinja**, a reference-based line art colorization method. MangaNinja
+automatically aligns the reference with the line art for colorization, demonstrating remarkable consistency. Additionally, users can achieve
+more complex tasks using point control. We hope that MangaNinja can accelerate the colorization process in the anime industry.
+
+![teaser](docs/teaser.gif)
+## 📢 News
+* 2025-01-15: Inference code and paper are released.
+* 2025-01-16: MangaNinja is available on windows, 6G VRAM need Auto install and Download Model. Thanks @sdbds ! You can found it [here](https://github.com/sdbds/MangaNinjia-for-windows). 🔥
+* 🏃: We will open an issue area to investigate user needs and adjust the model accordingly. This includes more memory-efficient structures, data formats for line art (such as binary line art), and considering retraining MangaNinjia on a better foundation model (sd3,flux).
+
+## 🛠️ Setup
+
+### 📦 Repository
+
+Clone the repository (requires git):
+
+```bash
+git clone https://github.com/ali-vilab/MangaNinjia.git
+cd MangaNinjia
+```
+
+### 💻 Dependencies
+
+Install with `conda`: 
+```bash
+conda env create -f environment.yaml
+conda activate MangaNinjia
+```
+### ⚙️ Weights
+* You could download them from HuggingFace: [StableDiffusion](https://modelscope.cn/models/AI-ModelScope/stable-diffusion-v1-5), [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14), [control_v11p_sd15_lineart](https://huggingface.co/lllyasviel/control_v11p_sd15_lineart) and [Annotators](https://huggingface.co/lllyasviel/Annotators/blob/main/sk_model.pth)
+* You could download our [MangaNinjia model](https://huggingface.co/Johanan0528/MangaNinjia) from HuggingFace 
+* The downloaded checkpoint directory has the following structure:
+```
+-- checkpoints
+    |-- StableDiffusion
+    |-- models
+        |-- clip-vit-large-patch14
+        |-- control_v11p_sd15_lineart
+        |-- Annotators
+            |--sk_model.pth
+    |-- MangaNinjia
+        |-- denoising_unet.pth
+        |-- reference_unet.pth
+        |-- point_net.pth
+        |-- controlnet.pth
+```
+
+
+## 🎮 Inference 
+```bash
+cd scripts
+bash infer.sh
+```
+
+You can find all results in `output/`. Enjoy!
+
+#### 📍 Inference settings
+
+The default settings are optimized for the best result. However, the behavior of the code can be customized:
+  - `--denoise_steps`: Number of denoising steps of each inference pass. For the original (DDIM) version, it's recommended to use 20-50 steps.
+  - `--is_lineart`: If the user provides an image and the task is to color the line art within that image, this parameter is not needed. However, if the input is already a line art and no additional extraction is necessary, then this parameter should be included.
+  - `--guidance_scale_ref`: Increasing makes the model more inclined to accept the guidance of the reference image.
+  - `--guidance_scale_point`: Increasing makes the model more inclined to input point guidance to achieve more customized colorization.
+  - `--point_ref_paths` and `--point_lineart_paths` (**optional**): Two 512x512 matrices are used to represent the matching points between the corresponding reference and line art with continuously increasing integers. That is, the coordinates of the matching points in both matrices will have the same values: 1, 2, 3, etc., while the values in other positions will be 0 (you can refer to the provided samples). Of course, we recommend using Gradio for point guidance.
+
+## 🌱 Gradio
+First, modify `./configs/inference.yaml` to set the path of model weight. Afterwards, run the script:
+```bash
+python run_gradio.py
+```
+The gradio demo would look like the UI shown below. 
+<table align="center">
+  <tr>
+    <td>
+      <img src="docs/gradio1.png" width="300" height="400">
+    </td>
+    <td>
+      <img src="docs/gradio2.png" width="300" height="400">
+    </td>
+  </tr>
+</table>
+A biref tutorial:
+
+1. Upload the reference image and target image. 
+
+    Note that for the target image, there are two modes: you can upload an RGB image, and the model will automatically extract the line art; or you can directly upload the line art by checking the 'input is lineart' option. 
+
+    The line art images are single-channel grayscale images, where the input consists of floating-point values with the background set to 0 and the line art close to 1. Additionally, we would like to further communicate with our users: if the line art you commonly use is binarized, please let us know. We will fine-tune the model and release an updated version to better suit your needs. 😆
+
+2. Click 'Process Images' to resize the images to 512*512 resolution.
+3. (Optional) **Starting from the reference image**, **alternately** click on the reference and target images in sequence to define matching points. Use 'Undo' to revert the last action.
+4. Click 'Generate' to produce the result.
+## 🌺 Acknowledgements
+This project is developped on the codebase of [MagicAnimate](https://github.com/magic-research/magic-animate). We appreciate this great work! 
+
+## 🎓 Citation
+
+Please cite our paper:
+
+```bibtex
+@article{liu2024manganinja,
+  author  = {Zhiheng Liu and Ka Leong Cheng and Xi Chen and Jie Xiao and Hao Ouyang and Kai Zhu and Yu Liu and Yujun Shen
+             and Qifeng Chen and Ping Luo},
+  title   = {MangaNinja: Line Art Colorization with Precise Reference Following},
+  journal = {CoRR},
+  volume  = {abs/xxxx.xxxxx},
+  year    = {2024}
+}
+```

configs/inference.yaml

@@ -0,0 +1,12 @@
+model_path:
+  pretrained_model_name_or_path: ./checkpoints/StableDiffusion
+  clip_vision_encoder_path: ./checkpoints/models/clip-vit-large-patch14
+  controlnet_model_name: './checkpoints/models/control_v11p_sd15_lineart'
+  annotator_ckpts_path: ./checkpoints/models/Annotators
+  manga_control_model_path: ./checkpoints/MangaNinjia/controlnet.pth
+  manga_reference_model_path: ./checkpoints/MangaNinjia/reference_unet.pth
+  manga_main_model_path: ./checkpoints/MangaNinjia/denoising_unet.pth
+  point_net_path: ./checkpoints/MangaNinjia/point_net.pth
+inference_config:
+  output_path: output
+  device: cuda

environment.yaml

@@ -0,0 +1,37 @@
+name: MangaNinjia
+channels:
+  - defaults
+dependencies:
+  - pip=24.2
+  - python=3.10.14
+  - pip:
+      - accelerate==0.31.0
+      - diffusers==0.27.2
+      - gradio==3.39.0
+      - gradio-client==1.3.0
+      - h5py==3.11.0
+      - huggingface-hub==0.24.6
+      - imageio==2.35.1
+      - imageio-ffmpeg==0.5.1
+      - importlib-metadata==8.4.0
+      - importlib-resources==6.4.5
+      - ipdb==0.13.13
+      - ipython==8.26.0
+      - ipywidgets==8.1.5
+      - kornia==0.7.3
+      - kornia-rs==0.1.5
+      - omegaconf==2.3.0
+      - opencv-python==4.10.0.84
+      - pandas==2.2.2
+      - pillow==10.4.0
+      - scikit-image==0.24.0
+      - scikit-learn==1.5.2
+      - scipy==1.14.1
+      - torch==2.3.0
+      - torchaudio==2.3.0
+      - torchmetrics==1.4.1
+      - torchvision==0.18.0
+      - tqdm==4.66.5
+      - transformers==4.44.1
+      - einops==0.8.0
+      - basicsr==1.3.5

infer.py

@@ -0,0 +1,277 @@
+import argparse
+import logging
+import os
+import random
+import numpy as np
+import torch
+from PIL import Image
+from tqdm.auto import tqdm
+from transformers import CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+import torch.nn as nn
+from inference.manganinjia_pipeline import MangaNinjiaPipeline
+from diffusers import (
+    ControlNetModel,
+    DiffusionPipeline,
+    DDIMScheduler,
+    AutoencoderKL,
+)
+from src.models.mutual_self_attention_multi_scale import ReferenceAttentionControl
+from src.models.unet_2d_condition import UNet2DConditionModel
+from src.models.refunet_2d_condition import RefUNet2DConditionModel
+from src.point_network import PointNet
+from src.annotator.lineart import BatchLineartDetector
+
+if "__main__" == __name__:
+    logging.basicConfig(level=logging.INFO)
+
+    # -------------------- Arguments --------------------
+    parser = argparse.ArgumentParser(
+        description="Run single-image MangaNinjia"
+    )
+    parser.add_argument(
+        "--output_dir", type=str, required=True, help="Output directory."
+    )
+
+    # inference setting
+    parser.add_argument(
+        "--denoise_steps",
+        type=int,
+        default=50,  # quantitative evaluation uses 50 steps
+        help="Diffusion denoising steps, more steps results in higher accuracy but slower inference speed.",
+    )
+
+    # resolution setting
+    parser.add_argument("--seed", type=int, default=None, help="Random seed.")
+
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--image_encoder_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--controlnet_model_name_or_path", type=str, required=True, help="Path to original controlnet."
+    )
+    parser.add_argument(
+        "--annotator_ckpts_path", type=str, required=True, help="Path to depth inpainting model."
+    )
+    parser.add_argument(
+        "--manga_reference_unet_path", type=str, required=True, help="Path to depth inpainting model."
+    )
+    parser.add_argument(
+        "--manga_main_model_path", type=str, required=True, help="Path to depth inpainting model."
+    )
+    parser.add_argument(
+        "--manga_controlnet_model_path", type=str, required=True, help="Path to depth inpainting model."
+    )
+    parser.add_argument(
+        "--point_net_path", type=str, required=True, help="Path to depth inpainting model."
+    )
+    parser.add_argument(
+        "--input_reference_paths",
+        nargs='+',
+        default=None,
+        help="input_image_paths",
+    )
+    parser.add_argument(
+        "--input_lineart_paths",
+        nargs='+',
+        default=None,
+        help="lineart_paths",
+    )
+    parser.add_argument(
+        "--point_ref_paths",
+        type=str,
+        default=None,
+        nargs="+",
+    )
+    parser.add_argument(
+        "--point_lineart_paths",
+        type=str,
+        default=None,
+        nargs="+",
+    )
+    parser.add_argument(
+        "--is_lineart",
+        action="store_true",
+        default=False
+    )
+    parser.add_argument(
+        "--guidance_scale_ref",
+        type=float,
+        default=1e-4,
+        help="guidance scale for reference image",
+    )
+    parser.add_argument(
+        "--guidance_scale_point",
+        type=float,
+        default=1e-4,
+        help="guidance scale for points",
+    )
+    args = parser.parse_args()
+    output_dir = args.output_dir
+    denoise_steps = args.denoise_steps
+    seed = args.seed
+    is_lineart = args.is_lineart
+    os.makedirs(output_dir, exist_ok=True)
+    logging.info(f"output dir = {output_dir}")
+    if args.input_reference_paths is not None:
+        assert len(args.input_reference_paths) == len(args.input_lineart_paths)
+    input_reference_paths = args.input_reference_paths
+    input_lineart_paths = args.input_lineart_paths
+    if args.point_ref_paths is not None:
+        point_ref_paths = args.point_ref_paths
+        point_lineart_paths = args.point_lineart_paths
+        assert len(point_ref_paths) == len(point_lineart_paths)
+    print(f"arguments: {args}")
+    if seed is None:
+        import time
+
+        seed = int(time.time())
+    generator = torch.cuda.manual_seed(seed)
+    # -------------------- Device --------------------
+    if torch.cuda.is_available():
+        device = torch.device("cuda")
+    else:
+        device = torch.device("cpu")
+        logging.warning("CUDA is not available. Running on CPU will be slow.")
+    logging.info(f"device = {device}")
+
+    # -------------------- Model --------------------
+    preprocessor = BatchLineartDetector(args.annotator_ckpts_path)
+    preprocessor.to(device,dtype=torch.float32) 
+    in_channels_reference_unet = 4
+    in_channels_denoising_unet = 4
+    in_channels_controlnet = 4
+    noise_scheduler = DDIMScheduler.from_pretrained(args.pretrained_model_name_or_path,subfolder='scheduler')
+    vae = AutoencoderKL.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder='vae'
+    )
+
+    denoising_unet = UNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path,subfolder="unet",
+        in_channels=in_channels_denoising_unet,
+        low_cpu_mem_usage=False,
+        ignore_mismatched_sizes=True
+    )
+            
+    reference_unet = RefUNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path,subfolder="unet",
+        in_channels=in_channels_reference_unet,
+        low_cpu_mem_usage=False,
+        ignore_mismatched_sizes=True
+    )
+    refnet_tokenizer = CLIPTokenizer.from_pretrained(args.image_encoder_path)
+    refnet_text_encoder = CLIPTextModel.from_pretrained(args.image_encoder_path)
+    refnet_image_encoder = CLIPVisionModelWithProjection.from_pretrained(args.image_encoder_path)
+        
+    controlnet = ControlNetModel.from_pretrained(
+        args.controlnet_model_name_or_path,
+        in_channels=in_channels_controlnet,
+        low_cpu_mem_usage=False,
+        ignore_mismatched_sizes=True
+    )
+    controlnet_tokenizer = CLIPTokenizer.from_pretrained(args.image_encoder_path)
+    controlnet_text_encoder = CLIPTextModel.from_pretrained(args.image_encoder_path)
+    controlnet_image_encoder = CLIPVisionModelWithProjection.from_pretrained(args.image_encoder_path)
+        
+
+    point_net=PointNet()
+
+
+
+    controlnet.load_state_dict(
+            torch.load(args.manga_controlnet_model_path, map_location="cpu"),
+            strict=False,
+            )
+    point_net.load_state_dict(
+            torch.load(args.point_net_path, map_location="cpu"),
+            strict=False,
+            )
+    reference_unet.load_state_dict(
+            torch.load(args.manga_reference_unet_path, map_location="cpu"),
+            strict=False,
+            )
+    denoising_unet.load_state_dict(
+            torch.load(args.manga_main_model_path, map_location="cpu"),
+            strict=False,
+            )
+    pipe = MangaNinjiaPipeline(
+            reference_unet=reference_unet,
+            controlnet=controlnet,
+            denoising_unet=denoising_unet,  
+            vae=vae,
+            refnet_tokenizer=refnet_tokenizer,
+            refnet_text_encoder=refnet_text_encoder,
+            refnet_image_encoder=refnet_image_encoder,
+            controlnet_tokenizer=controlnet_tokenizer,
+            controlnet_text_encoder=controlnet_text_encoder,
+            controlnet_image_encoder=controlnet_image_encoder,
+            scheduler=noise_scheduler,
+            point_net=point_net
+        )
+    pipe = pipe.to(torch.device(device))
+
+    # -------------------- Inference and saving --------------------
+    with torch.no_grad():
+        for i in range(len(input_reference_paths)):
+            input_reference_path = input_reference_paths[i]
+            input_lineart_path = input_lineart_paths[i]
+
+            # save path
+            rgb_name_base = os.path.splitext(os.path.basename(input_reference_path))[0]
+            pred_name_base = rgb_name_base + "_colorized"
+            lineart_name_base = rgb_name_base + "_lineart"
+            colored_save_path = os.path.join(
+                output_dir, f"{pred_name_base}.png"
+            )
+            lineart_save_path = os.path.join(
+                output_dir, f"{lineart_name_base}.png"
+            )
+            if point_ref_paths is not None:
+                point_ref_path = point_ref_paths[i]
+                point_lineart_path = point_lineart_paths[i]
+                point_ref = torch.from_numpy(np.load(point_ref_path)).unsqueeze(0).unsqueeze(0)
+                point_main = torch.from_numpy(np.load(point_lineart_path)).unsqueeze(0).unsqueeze(0)
+            else:
+                matrix1 = np.zeros((512, 512), dtype=np.uint8)
+                matrix2 = np.zeros((512, 512), dtype=np.uint8)
+                point_ref = torch.from_numpy(matrix1).unsqueeze(0).unsqueeze(0)
+                point_main = torch.from_numpy(matrix2).unsqueeze(0).unsqueeze(0)
+            ref_image = Image.open(input_reference_path)
+            ref_image = ref_image.resize((512, 512))
+            target_image = Image.open(input_lineart_path)
+            target_image = target_image.resize((512, 512))
+            pipe_out = pipe(
+                is_lineart,
+                ref_image,
+                target_image,
+                target_image,
+                denosing_steps=denoise_steps,
+                processing_res=512,
+                match_input_res=True,
+                batch_size=1,
+                show_progress_bar=True,
+                guidance_scale_ref=args.guidance_scale_ref,
+                guidance_scale_point=args.guidance_scale_point,
+                preprocessor=preprocessor,
+                generator=generator,
+                point_ref=point_ref,  
+                point_main=point_main,  
+            )
+
+            if os.path.exists(colored_save_path):
+                logging.warning(f"Existing file: '{colored_save_path}' will be overwritten")
+            image = pipe_out.img_pil
+            lineart = pipe_out.to_save_dict['edge2_black']
+            image.save(colored_save_path)
+            lineart.save(lineart_save_path)

inference/manganinjia_pipeline.py

@@ -0,0 +1,486 @@
+
+from typing import Any, Dict, Union
+import torchvision.transforms as transforms
+import torch
+from torch.utils.data import DataLoader, TensorDataset
+import numpy as np
+from tqdm.auto import tqdm
+from PIL import Image
+from diffusers import (
+    DiffusionPipeline,
+    ControlNetModel,
+    DDIMScheduler,
+    AutoencoderKL,
+)
+from diffusers.utils import BaseOutput
+from transformers import CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor
+from transformers import CLIPVisionModelWithProjection
+
+from utils.image_util import resize_max_res,chw2hwc
+from src.point_network import PointNet
+from src.models.mutual_self_attention_multi_scale import ReferenceAttentionControl
+from src.models.unet_2d_condition import UNet2DConditionModel
+from src.models.refunet_2d_condition import RefUNet2DConditionModel
+
+
+class MangaNinjiaPipelineOutput(BaseOutput):
+    img_np: np.ndarray
+    img_pil: Image.Image
+    to_save_dict: dict
+
+
+class MangaNinjiaPipeline(DiffusionPipeline):
+    rgb_latent_scale_factor = 0.18215
+    
+    def __init__(self,
+        reference_unet: RefUNet2DConditionModel,
+        controlnet: ControlNetModel,
+        denoising_unet: UNet2DConditionModel,
+        vae: AutoencoderKL,
+        refnet_tokenizer: CLIPTokenizer,
+        refnet_text_encoder: CLIPTextModel,
+        refnet_image_encoder: CLIPVisionModelWithProjection,
+        controlnet_tokenizer: CLIPTokenizer,
+        controlnet_text_encoder: CLIPTextModel,
+        controlnet_image_encoder: CLIPVisionModelWithProjection,
+        scheduler: DDIMScheduler,
+        point_net: PointNet
+    ):
+        super().__init__()
+            
+        self.register_modules(
+            reference_unet=reference_unet,
+            controlnet=controlnet,
+            denoising_unet=denoising_unet,       
+            vae=vae,
+            refnet_tokenizer=refnet_tokenizer,
+            refnet_text_encoder=refnet_text_encoder,
+            refnet_image_encoder=refnet_image_encoder,
+            controlnet_tokenizer=controlnet_tokenizer,
+            controlnet_text_encoder=controlnet_text_encoder,
+            controlnet_image_encoder=controlnet_image_encoder,
+            point_net=point_net,
+            scheduler=scheduler,
+        )
+        self.empty_text_embed = None
+        self.clip_image_processor = CLIPImageProcessor()
+        
+    @torch.no_grad()
+    def __call__(
+        self,
+        is_lineart: bool,
+        ref1: Image.Image,
+        raw2: Image.Image,
+        edit2: Image.Image,
+        denosing_steps: int = 20,
+        processing_res: int = 512,
+        match_input_res: bool = True,
+        batch_size: int = 0,
+        show_progress_bar: bool = True,
+        guidance_scale_ref: float = 7,
+        guidance_scale_point: float = 12,
+        preprocessor=None,
+        generator=None,
+        point_ref=None,
+        point_main=None,
+    ) -> MangaNinjiaPipelineOutput:
+
+        device = self.device
+        
+        input_size = raw2.size
+        point_ref=point_ref.float().to(device)
+        point_main=point_main.float().to(device)
+        def img2embeds(img, image_enc):
+            clip_image = self.clip_image_processor.preprocess(
+                img, return_tensors="pt"
+            ).pixel_values
+            clip_image_embeds = image_enc(
+                clip_image.to(device, dtype=image_enc.dtype)
+            ).image_embeds
+            encoder_hidden_states = clip_image_embeds.unsqueeze(1)
+            return encoder_hidden_states
+        if self.reference_unet:
+            refnet_encoder_hidden_states = img2embeds(ref1, self.refnet_image_encoder)
+        else:
+            refnet_encoder_hidden_states = None
+        if self.controlnet:
+            controlnet_encoder_hidden_states = img2embeds(ref1, self.controlnet_image_encoder)
+        else:
+            controlnet_encoder_hidden_states = None
+
+        prompt = ""
+        def prompt2embeds(prompt, tokenizer, text_encoder):
+            text_inputs = tokenizer(
+                prompt,
+                padding="do_not_pad",
+                max_length=tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids.to(device) #[1,2]
+            empty_text_embed = text_encoder(text_input_ids)[0].to(self.dtype)
+            uncond_encoder_hidden_states = empty_text_embed.repeat((1, 1, 1))[:,0,:].unsqueeze(0)
+            return uncond_encoder_hidden_states
+        if self.reference_unet:
+            refnet_uncond_encoder_hidden_states = prompt2embeds(prompt, self.refnet_tokenizer, self.refnet_text_encoder)
+        else:
+            refnet_uncond_encoder_hidden_states = None
+        if self.controlnet:
+            controlnet_uncond_encoder_hidden_states = prompt2embeds(prompt, self.controlnet_tokenizer, self.controlnet_text_encoder)
+        else:
+            controlnet_uncond_encoder_hidden_states = None
+
+        do_classifier_free_guidance = guidance_scale_ref > 1.0
+        
+        # adjust the input resolution.
+        if not match_input_res:
+            assert (
+                processing_res is not None                
+            )," Value Error: `resize_output_back` is only valid with "
+        
+        assert processing_res >= 0
+        assert denosing_steps >= 1
+        
+        # --------------- Image Processing ------------------------
+        # Resize image
+        if processing_res > 0:
+            def resize_img(img):
+                img = resize_max_res(img, max_edge_resolution=processing_res)
+                return img
+            ref1 = resize_img(ref1)
+            raw2 = resize_img(raw2)
+            edit2 = resize_img(edit2)
+        
+        # Normalize image
+        def normalize_img(img):
+            img = img.convert("RGB")
+            img = np.array(img)
+
+            # Normalize RGB Values.
+            rgb = np.transpose(img,(2,0,1))
+            rgb_norm = rgb / 255.0 * 2.0 - 1.0
+            rgb_norm = torch.from_numpy(rgb_norm).to(self.dtype)
+            rgb_norm = rgb_norm.to(device)
+            img = rgb_norm
+            assert img.min() >= -1.0 and img.max() <= 1.0
+            return img
+        raw2_real = raw2.convert('L')
+        ref1 = normalize_img(ref1)
+        raw2 = normalize_img(raw2)
+        edit2 = normalize_img(edit2)
+        single_rgb_dataset = TensorDataset(ref1[None], raw2[None], edit2[None])
+
+        
+        # find the batch size
+        if batch_size>0:
+            _bs = batch_size
+        else:
+            _bs = 1
+        point_ref=self.point_net(point_ref)
+        point_main=self.point_net(point_main)
+        single_rgb_loader = DataLoader(single_rgb_dataset,batch_size=_bs,shuffle=False)
+
+        # classifier guidance
+        if do_classifier_free_guidance:
+            if self.reference_unet:
+                refnet_encoder_hidden_states = torch.cat(
+                    [refnet_uncond_encoder_hidden_states, refnet_encoder_hidden_states,refnet_encoder_hidden_states], dim=0
+                )
+            else:
+                refnet_encoder_hidden_states = None
+
+            if self.controlnet:
+                controlnet_encoder_hidden_states = torch.cat(
+                    [controlnet_uncond_encoder_hidden_states, controlnet_encoder_hidden_states,controlnet_encoder_hidden_states], dim=0
+                )
+            else:
+                controlnet_encoder_hidden_states = None
+
+        if self.reference_unet:
+            reference_control_writer = ReferenceAttentionControl(
+                self.reference_unet,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                mode="write",
+                batch_size=batch_size,
+                fusion_blocks="full",
+            )
+            reference_control_reader = ReferenceAttentionControl(
+                self.denoising_unet,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                mode="read",
+                batch_size=batch_size,
+                fusion_blocks="full",
+            )
+        else:
+            reference_control_writer = None
+            reference_control_reader = None
+            
+        if show_progress_bar:
+            iterable_bar = tqdm(
+                single_rgb_loader, desc=" " * 2 + "Inference batches", leave=False
+            )
+        else:
+            iterable_bar = single_rgb_loader
+        
+        assert len(iterable_bar) == 1
+        for batch in iterable_bar:
+            (ref1, raw2, edit2) = batch  # here the image is still around 0-1
+            if is_lineart:
+                raw2 = raw2_real
+            img_pred, to_save_dict = self.single_infer(
+                is_lineart=is_lineart,
+                ref1=ref1,
+                raw2=raw2,
+                edit2=edit2,
+                num_inference_steps=denosing_steps,
+                show_pbar=show_progress_bar,
+                guidance_scale_ref=guidance_scale_ref,
+                guidance_scale_point=guidance_scale_point,
+                refnet_encoder_hidden_states=refnet_encoder_hidden_states,
+                controlnet_encoder_hidden_states=controlnet_encoder_hidden_states,
+                reference_control_writer=reference_control_writer,
+                reference_control_reader=reference_control_reader,
+                preprocessor=preprocessor,
+                generator=generator,
+                point_ref=point_ref,
+                point_main=point_main
+            )
+            for k, v in to_save_dict.items():
+                if k =='edge2_black':
+                    to_save_dict[k] = Image.fromarray(
+                   ((to_save_dict['edge2_black'][:,0].squeeze().detach().cpu().numpy() + 1.) / 2 * 255).astype(np.uint8)
+                    )
+                else:
+                    try:
+                        to_save_dict[k] = Image.fromarray(
+                        chw2hwc(((v.squeeze().detach().cpu().numpy() + 1.) / 2 * 255).astype(np.uint8))
+                        )
+                    except:
+                        import ipdb;ipdb.set_trace()
+        
+        torch.cuda.empty_cache()  # clear vram cache for ensembling
+        
+        # ----------------- Post processing -----------------        
+        # Convert to numpy
+        img_pred = img_pred.squeeze().cpu().numpy().astype(np.float32)
+        img_pred_np = (((img_pred + 1.) / 2.) * 255).astype(np.uint8)
+        img_pred_np = chw2hwc(img_pred_np)
+        img_pred_pil = Image.fromarray(img_pred_np)
+
+        # Resize back to original resolution
+        if match_input_res:
+            img_pred_pil = img_pred_pil.resize(input_size)
+            img_pred_np = np.asarray(img_pred_pil)        
+
+        return MangaNinjiaPipelineOutput(
+            img_np=img_pred_np,
+            img_pil=img_pred_pil,
+            to_save_dict=to_save_dict
+        )
+
+    
+    def __encode_empty_text(self):
+        """
+        Encode text embedding for empty prompt
+        """
+        prompt = ""
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="do_not_pad",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids.to(self.text_encoder.device) #[1,2]
+        # print(text_input_ids.shape)
+        self.empty_text_embed = self.text_encoder(text_input_ids)[0].to(self.dtype) #[1,2,1024]
+
+    def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
+        # get the original timestep using init_timestep
+        if denoising_start is None:
+            init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+            t_start = max(num_inference_steps - init_timestep, 0)
+        else:
+            t_start = 0
+
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+        # Strength is irrelevant if we directly request a timestep to start at;
+        # that is, strength is determined by the denoising_start instead.
+        if denoising_start is not None:
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (denoising_start * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            timesteps = list(filter(lambda ts: ts < discrete_timestep_cutoff, timesteps))
+            return torch.tensor(timesteps), len(timesteps)
+
+        return timesteps, num_inference_steps - t_start
+        
+    @torch.no_grad()
+    def single_infer(
+        self,
+        is_lineart: bool,
+        ref1: torch.Tensor,
+        raw2: torch.Tensor,
+        edit2: torch.Tensor,
+        num_inference_steps: int,
+        show_pbar: bool,
+        guidance_scale_ref: float,
+        guidance_scale_point: float,
+        refnet_encoder_hidden_states: torch.Tensor,
+        controlnet_encoder_hidden_states: torch.Tensor,
+        reference_control_writer: ReferenceAttentionControl,
+        reference_control_reader: ReferenceAttentionControl,
+        preprocessor,
+        generator,
+        point_ref,
+        point_main
+    ):
+        do_classifier_free_guidance = guidance_scale_ref > 1.0
+        device = ref1.device
+        to_save_dict = {
+            'ref1': ref1,
+        }
+        
+        # Set timesteps: inherit from the diffuison pipeline
+        self.scheduler.set_timesteps(num_inference_steps, device=device) # here the numbers of the steps is only 10.
+        timesteps = self.scheduler.timesteps  # [T]
+        
+        # encode image
+        ref1_latents = self.encode_RGB(ref1, generator=generator) # 1/8 Resolution with a channel nums of 4. 
+        edge2_src = raw2
+
+        timesteps_add,_=self.get_timesteps(num_inference_steps, 1.0, device, denoising_start=None)
+        if is_lineart is not True:
+            edge2 = preprocessor(edge2_src)
+        else:
+            gray_image_np = np.array(edge2_src)
+            gray_image_np = gray_image_np / 255.0
+            edge2 = torch.from_numpy(gray_image_np.astype(np.float32)).unsqueeze(0).unsqueeze(0).cuda()
+        edge2[edge2<=0.24]=0
+        edge2_black = edge2.repeat(1, 3, 1, 1) * 2 - 1.
+        to_save_dict['edge2_black']=edge2_black
+
+        edge2 = edge2.repeat(1, 3, 1, 1) * 2 - 1.
+        to_save_dict['edge2'] = (1-((edge2+1.)/2))*2-1
+        
+        noisy_edit2_latents = torch.randn(
+            ref1_latents.shape, device=device, dtype=self.dtype
+        )  # [B, 4, H/8, W/8]
+            
+   
+        # Denoising loop
+        if show_pbar:
+            iterable = tqdm(
+                enumerate(timesteps),
+                total=len(timesteps),
+                leave=False,
+                desc=" " * 4 + "Diffusion denoising",
+            )
+        else:
+            iterable = enumerate(timesteps)
+
+        for i, t in iterable:
+            
+            refnet_input = ref1_latents
+            controlnet_inputs = (noisy_edit2_latents, edge2)
+            unet_input = torch.cat([noisy_edit2_latents], dim=1)
+
+            if i == 0:
+                if self.reference_unet:
+                    self.reference_unet(
+                        refnet_input.repeat(
+                            (3 if do_classifier_free_guidance else 1), 1, 1, 1
+                        ),
+                        torch.zeros_like(t),
+                        
+                        encoder_hidden_states=refnet_encoder_hidden_states,
+                        return_dict=False,
+                    )
+                    reference_control_reader.update(reference_control_writer,point_embedding_ref=point_ref,point_embedding_main=point_main)#size不对
+
+            if self.controlnet:
+                noisy_latents, controlnet_cond = controlnet_inputs
+                down_block_res_samples, mid_block_res_sample = self.controlnet(
+                    noisy_latents.repeat(
+                        (3 if do_classifier_free_guidance else 1), 1, 1, 1
+                    ),
+                    t,
+                    encoder_hidden_states=controlnet_encoder_hidden_states,
+                    controlnet_cond=controlnet_cond.repeat(
+                        (3 if do_classifier_free_guidance else 1), 1, 1, 1
+                    ),
+                    return_dict=False,
+                )
+            else:
+                down_block_res_samples, mid_block_res_sample = None, None
+
+            # predict the noise residual
+            noise_pred = self.denoising_unet(
+                unet_input.repeat(
+                    (3 if do_classifier_free_guidance else 1), 1, 1, 1
+                ).to(dtype=self.denoising_unet.dtype), 
+                t, 
+                encoder_hidden_states=refnet_encoder_hidden_states,
+                down_block_additional_residuals=down_block_res_samples,
+                mid_block_additional_residual=mid_block_res_sample,
+            ).sample  # [B, 4, h, w]
+            noise_pred_uncond, noise_pred_ref, noise_pred_point = noise_pred.chunk(3)
+            noise_pred_1 = noise_pred_uncond + guidance_scale_ref * (
+                noise_pred_ref - noise_pred_uncond
+            )
+            noise_pred_2 = noise_pred_ref + guidance_scale_point * (
+                noise_pred_point - noise_pred_ref
+            )
+            noise_pred=(noise_pred_1+noise_pred_2)/2
+            noisy_edit2_latents = self.scheduler.step(noise_pred, t, noisy_edit2_latents).prev_sample
+        
+        reference_control_reader.clear()
+        reference_control_writer.clear()
+        torch.cuda.empty_cache()
+
+        # clip prediction
+        edit2 = self.decode_RGB(noisy_edit2_latents)
+        edit2 = torch.clip(edit2, -1.0, 1.0)
+
+        return edit2, to_save_dict
+        
+    
+    def encode_RGB(self, rgb_in: torch.Tensor, generator) -> torch.Tensor:
+        """
+        Encode RGB image into latent.
+
+        Args:
+            rgb_in (`torch.Tensor`):
+                Input RGB image to be encoded.
+
+        Returns:
+            `torch.Tensor`: Image latent.
+        """
+        
+        # generator = None
+        rgb_latent = self.vae.encode(rgb_in).latent_dist.sample(generator)
+        rgb_latent = rgb_latent * self.rgb_latent_scale_factor
+        return rgb_latent
+    
+    def decode_RGB(self, rgb_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode depth latent into depth map.
+
+        Args:
+            rgb_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded depth map.
+        """
+
+        rgb_latent = rgb_latent / self.rgb_latent_scale_factor
+        rgb_out = self.vae.decode(rgb_latent, return_dict=False)[0]
+        return rgb_out
+
+

requirements.txt

@@ -0,0 +1,30 @@
+accelerate==0.31.0
+diffusers==0.27.2
+gradio==3.39.0
+gradio-client==1.3.0
+h5py==3.11.0
+huggingface-hub==0.24.6
+imageio==2.35.1
+imageio-ffmpeg==0.5.1
+importlib-metadata==8.4.0
+importlib-resources==6.4.5
+ipdb==0.13.13
+ipython==8.26.0
+ipywidgets==8.1.5
+kornia==0.7.3
+kornia-rs==0.1.5
+omegaconf==2.3.0
+opencv-python==4.10.0.84
+pandas==2.2.2
+pillow==10.4.0
+scikit-image==0.24.0
+scikit-learn==1.5.2
+scipy==1.14.1
+torch==2.3.0
+torchaudio==2.3.0
+torchmetrics==1.4.1
+torchvision==0.18.0
+tqdm==4.66.5
+transformers==4.44.1
+einops==0.8.0
+basicsr==1.3.5

run_gradio.py

@@ -0,0 +1,381 @@
+import gradio as gr
+import numpy as np
+from PIL import Image, ImageDraw
+import cv2
+import gradio as gr
+import torch
+import torch.nn.functional as F
+from omegaconf import OmegaConf
+import numpy as np
+import os
+import re
+from PIL import Image, ImageDraw
+import cv2
+#
+from transformers import CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+import torch.nn as nn
+from inference.manganinjia_pipeline import MangaNinjiaPipeline
+from diffusers import (
+    ControlNetModel,
+    DiffusionPipeline,
+    DDIMScheduler,
+    AutoencoderKL,
+)
+from src.models.mutual_self_attention_multi_scale import ReferenceAttentionControl
+from src.models.unet_2d_condition import UNet2DConditionModel
+from src.models.refunet_2d_condition import RefUNet2DConditionModel
+from src.point_network import PointNet
+from src.annotator.lineart import BatchLineartDetector
+val_configs = OmegaConf.load('./configs/inference.yaml')
+# === load the checkpoint ===
+pretrained_model_name_or_path = val_configs.model_path.pretrained_model_name_or_path
+refnet_clip_vision_encoder_path = val_configs.model_path.clip_vision_encoder_path
+controlnet_clip_vision_encoder_path = val_configs.model_path.clip_vision_encoder_path
+controlnet_model_name_or_path = val_configs.model_path.controlnet_model_name
+annotator_ckpts_path = val_configs.model_path.annotator_ckpts_path
+
+output_root = val_configs.inference_config.output_path
+device = val_configs.inference_config.device
+preprocessor = BatchLineartDetector(annotator_ckpts_path)
+in_channels_reference_unet = 4
+in_channels_denoising_unet = 4
+in_channels_controlnet = 4
+noise_scheduler = DDIMScheduler.from_pretrained(pretrained_model_name_or_path,subfolder='scheduler')
+vae = AutoencoderKL.from_pretrained(
+    pretrained_model_name_or_path,
+    subfolder='vae'
+)
+
+denoising_unet = UNet2DConditionModel.from_pretrained(
+    pretrained_model_name_or_path,subfolder="unet",
+    in_channels=in_channels_denoising_unet,
+    low_cpu_mem_usage=False,
+    ignore_mismatched_sizes=True
+)
+        
+reference_unet = RefUNet2DConditionModel.from_pretrained(
+    pretrained_model_name_or_path,subfolder="unet",
+    in_channels=in_channels_reference_unet,
+    low_cpu_mem_usage=False,
+    ignore_mismatched_sizes=True
+)
+refnet_tokenizer = CLIPTokenizer.from_pretrained(refnet_clip_vision_encoder_path)
+refnet_text_encoder = CLIPTextModel.from_pretrained(refnet_clip_vision_encoder_path)
+refnet_image_enc = CLIPVisionModelWithProjection.from_pretrained(refnet_clip_vision_encoder_path)
+       
+controlnet = ControlNetModel.from_pretrained(
+    controlnet_model_name_or_path,
+    in_channels=in_channels_controlnet,
+    low_cpu_mem_usage=False,
+    ignore_mismatched_sizes=True
+)
+controlnet_tokenizer = CLIPTokenizer.from_pretrained(controlnet_clip_vision_encoder_path)
+controlnet_text_encoder = CLIPTextModel.from_pretrained(controlnet_clip_vision_encoder_path)
+controlnet_image_enc = CLIPVisionModelWithProjection.from_pretrained(controlnet_clip_vision_encoder_path)
+      
+
+point_net=PointNet()
+reference_control_writer = ReferenceAttentionControl(
+            reference_unet,
+            do_classifier_free_guidance=False,
+            mode="write",
+            fusion_blocks="full",
+        )
+reference_control_reader = ReferenceAttentionControl(
+    denoising_unet,
+    do_classifier_free_guidance=False,
+    mode="read",
+    fusion_blocks="full",
+)
+
+
+
+controlnet.load_state_dict(
+        torch.load(val_configs.model_path.manga_control_model_path, map_location="cpu"),
+        strict=False,
+        )
+point_net.load_state_dict(
+        torch.load(val_configs.model_path.point_net_path, map_location="cpu"),
+        strict=False,
+        )
+reference_unet.load_state_dict(
+        torch.load(val_configs.model_path.manga_reference_model_path, map_location="cpu"),
+        strict=False,
+        )
+denoising_unet.load_state_dict(
+        torch.load(val_configs.model_path.manga_main_model_path, map_location="cpu"),
+        strict=False,
+        )
+pipe = MangaNinjiaPipeline(
+        reference_unet=reference_unet,
+        controlnet=controlnet,
+        denoising_unet=denoising_unet,  
+        vae=vae,
+        refnet_tokenizer=refnet_tokenizer,
+        refnet_text_encoder=refnet_text_encoder,
+        refnet_image_encoder=refnet_image_enc,
+        controlnet_tokenizer=controlnet_tokenizer,
+        controlnet_text_encoder=controlnet_text_encoder,
+        controlnet_image_encoder=controlnet_image_enc,
+        scheduler=noise_scheduler,
+        point_net=point_net
+    )
+pipe = pipe.to(torch.device(device))
+def string_to_np_array(coord_string):
+    coord_string = coord_string.strip('[]')
+    coords = re.findall(r'\d+', coord_string)
+    coords = list(map(int, coords))
+    coord_array = np.array(coords).reshape(-1, 2)
+    return coord_array
+def infer_single(is_lineart, ref_image, target_image, output_coords_ref, output_coords_base, seed = -1, num_inference_steps=20, guidance_scale_ref = 9, guidance_scale_point =15 ):
+    """
+    mask: 0/1 1-channel  np.array
+    image: rgb           np.array
+    """
+    generator = torch.cuda.manual_seed(seed)
+    matrix1 = np.zeros((512, 512), dtype=np.uint8)
+    matrix2 = np.zeros((512, 512), dtype=np.uint8)
+    output_coords_ref = string_to_np_array(output_coords_ref)
+    output_coords_base = string_to_np_array(output_coords_base)
+    for index, (coords_ref,coords_base) in enumerate(zip(output_coords_ref,output_coords_base)):
+        y1, x1 = coords_ref 
+        y2, x2 = coords_base
+        matrix1[y1, x1] = index + 1
+        matrix2[y2, x2] = index + 1
+    point_ref = torch.from_numpy(matrix1).unsqueeze(0).unsqueeze(0)
+    point_main = torch.from_numpy(matrix2).unsqueeze(0).unsqueeze(0)
+    preprocessor.to(device,dtype=torch.float32) 
+    pipe_out = pipe(
+        is_lineart,
+        ref_image,
+        target_image,
+        target_image,
+        denosing_steps=num_inference_steps,
+        processing_res=512,
+        match_input_res=True,
+        batch_size=1,
+        show_progress_bar=True,
+        guidance_scale_ref=guidance_scale_ref,
+        guidance_scale_point=guidance_scale_point,
+        preprocessor=preprocessor,
+        generator=generator,
+        point_ref=point_ref,  
+        point_main=point_main,  
+    )
+    return pipe_out
+
+
+def inference_single_image(ref_image, 
+                           tar_image, 
+                           ddim_steps, 
+                           scale_ref,
+                           scale_point, 
+                           seed,
+                           output_coords1,
+                           output_coords2,
+                           is_lineart
+                           ):
+    if seed == -1:
+        seed = np.random.randint(10000)
+    pipe_out = infer_single(is_lineart, ref_image, tar_image, output_coords_ref=output_coords1, output_coords_base=output_coords2,seed=seed ,num_inference_steps=ddim_steps, guidance_scale_ref = scale_ref, guidance_scale_point = scale_point
+                           )
+    return pipe_out
+clicked_points_img1 = []
+clicked_points_img2 = []
+current_img_idx = 0  
+max_clicks = 14  
+point_size = 8  
+colors = [(255, 0, 0), (0, 255, 0)] 
+
+# Process images: resizing them to 512x512
+def process_image(ref, base):
+    ref_resized = cv2.resize(ref, (512, 512))  # Note OpenCV resize order is (width, height)
+    base_resized = cv2.resize(base, (512, 512))
+    return ref_resized, base_resized
+
+# Convert string to numpy array of coordinates
+def string_to_np_array(coord_string):
+    coord_string = coord_string.strip('[]')
+    coords = re.findall(r'\d+', coord_string)
+    coords = list(map(int, coords))
+    coord_array = np.array(coords).reshape(-1, 2)
+    return coord_array
+
+# Function to handle click events
+def get_select_coords(img1, img2, evt: gr.SelectData):
+    global clicked_points_img1, clicked_points_img2, current_img_idx
+    click_coords = (evt.index[1], evt.index[0])
+
+    if current_img_idx == 0:
+        clicked_points_img1.append(click_coords)
+        if len(clicked_points_img1) > max_clicks:
+            clicked_points_img1 = []
+        current_img = img1
+        clicked_points = clicked_points_img1
+    else:
+        clicked_points_img2.append(click_coords)
+        if len(clicked_points_img2) > max_clicks:
+            clicked_points_img2 = []
+        current_img = img2
+        clicked_points = clicked_points_img2
+
+    current_img_idx = 1 - current_img_idx
+    img_pil = Image.fromarray(current_img.astype('uint8'))
+    draw = ImageDraw.Draw(img_pil)
+    for idx, point in enumerate(clicked_points):
+        x, y = point
+        color = colors[current_img_idx]
+        for dx in range(-point_size, point_size + 1):
+            for dy in range(-point_size, point_size + 1):
+                if 0 <= y + dy < img_pil.size[0] and 0 <= x + dx < img_pil.size[1]:
+                    draw.point((y+dy, x+dx), fill=color)
+
+    img_out = np.array(img_pil)
+    coord_array = np.array([(x, y) for x, y in clicked_points])
+    return img_out, coord_array
+
+# Function to clear the clicked points
+def undo_last_point(ref, base):
+    global clicked_points_img1, clicked_points_img2, current_img_idx
+    current_img_idx=1-current_img_idx
+    if current_img_idx == 0 and clicked_points_img1:
+        clicked_points_img1.pop()  # Undo last point in ref
+    elif current_img_idx == 1 and clicked_points_img2:
+        clicked_points_img2.pop()  # Undo last point in base
+
+    # After removing the last point, redraw the image without it
+    if current_img_idx == 0:
+        current_img = ref
+        current_img_other = base
+        clicked_points = clicked_points_img1
+        clicked_points_other = clicked_points_img2
+    else:
+        current_img = base
+        current_img_other = ref
+        clicked_points = clicked_points_img2
+        clicked_points_other = clicked_points_img1
+
+    # Redraw the image without the last point
+    img_pil = Image.fromarray(current_img.astype('uint8'))
+    draw = ImageDraw.Draw(img_pil)
+    for idx, point in enumerate(clicked_points):
+        x, y = point
+        color = colors[current_img_idx]
+        for dx in range(-point_size, point_size + 1):
+            for dy in range(-point_size, point_size + 1):
+                if 0 <= y + dy < img_pil.size[0] and 0 <= x + dx < img_pil.size[1]:
+                    draw.point((y+dy, x+dx), fill=color)
+    img_out = np.array(img_pil)
+
+
+    img_pil_other = Image.fromarray(current_img_other.astype('uint8'),)
+    draw_other = ImageDraw.Draw(img_pil_other)
+    for idx, point in enumerate(clicked_points_other):
+        x, y = point
+        color = colors[1-current_img_idx]
+        for dx in range(-point_size, point_size + 1):
+            for dy in range(-point_size, point_size + 1):
+                if 0 <= y + dy < img_pil.size[0] and 0 <= x + dx < img_pil.size[1]:
+                    draw_other.point((y+dy, x+dx), fill=color)
+    img_out_other = np.array(img_pil_other)
+
+    coord_array = np.array([(x, y) for x, y in clicked_points])
+    # Return the updated image and coordinates as text
+    updated_coords = str(coord_array.tolist())
+    
+    # If current_img_idx is 0, it means we are working with ref, so return for ref
+    if current_img_idx == 0:
+        coord_array2 = np.array([(x, y) for x, y in clicked_points_img2])
+        updated_coords2 = str(coord_array2.tolist())
+        return img_out, updated_coords, img_out_other, updated_coords2  # for ref image
+    else:
+        coord_array1 = np.array([(x, y) for x, y in clicked_points_img1])
+        updated_coords1 = str(coord_array1.tolist())
+        return img_out_other, updated_coords1, img_out, updated_coords  # for base image
+
+
+# Main function to run the image processing
+def run_local(ref, base, *args):
+    image = Image.fromarray(base)
+    ref_image = Image.fromarray(ref)
+    
+    pipe_out = inference_single_image(ref_image.copy(), image.copy(), *args)
+    to_save_dict = pipe_out.to_save_dict
+    to_save_dict['edit2'] = pipe_out.img_pil
+    return [to_save_dict['edit2'], to_save_dict['edge2_black']]
+
+with gr.Blocks() as demo:
+    with gr.Column():
+        gr.Markdown("#  MangaNinja: Line Art Colorization with Precise Reference Following")
+        
+        with gr.Row():
+            baseline_gallery = gr.Gallery(label='Output', show_label=True, elem_id="gallery", columns=1, height=768)
+            
+            with gr.Accordion("Advanced Option", open=True):
+                num_samples = 1
+                ddim_steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=50, step=1)
+                scale_ref = gr.Slider(label="Guidance of ref", minimum=0, maximum=30.0, value=9, step=0.1)
+                scale_point = gr.Slider(label="Guidance of points", minimum=0, maximum=30.0, value=15, step=0.1)
+                is_lineart = gr.Checkbox(label="Input is lineart", value=False)
+                seed = gr.Slider(label="Seed", minimum=-1, maximum=999999999, step=1, value=-1)
+                
+                gr.Markdown("### Tutorial")
+                gr.Markdown("1. Upload the reference image and target image. Note that for the target image, there are two modes: you can upload an RGB image, and the model will automatically extract the line art; or you can directly upload the line art by checking the 'input is lineart' option.")
+                gr.Markdown("2. Click 'Process Images' to resize the images to 512*512 resolution.")
+                gr.Markdown("3. (Optional) **Starting from the reference image**, **alternately** click on the reference and target images in sequence to define matching points. Use 'Undo' to revert the last action.")
+                gr.Markdown("4. Click 'Generate' to produce the result.")
+        gr.Markdown("# Upload the reference image and target image")
+        
+        with gr.Row():
+            ref = gr.Image(label="Reference Image",)
+            base = gr.Image(label="Target Image",)
+        gr.Button("Process Images").click(process_image, inputs=[ref, base], outputs=[ref, base])
+            
+        with gr.Row():
+            output_img1 = gr.Image(label="Reference Output")
+            output_coords1 = gr.Textbox(lines=2, label="Clicked Coordinates Image 1 (npy format)")
+            output_img2 = gr.Image(label="Base Output")
+            output_coords2 = gr.Textbox(lines=2, label="Clicked Coordinates Image 2 (npy format)")
+
+        # Image click select functions
+        ref.select(get_select_coords, [ref, base], [output_img1, output_coords1])
+        base.select(get_select_coords, [ref, base], [output_img2, output_coords2])
+        
+        # Undo button
+        undo_button = gr.Button("Undo")
+        undo_button.click(undo_last_point, inputs=[ref, base], outputs=[output_img1, output_coords1, output_img2, output_coords2])
+
+        run_local_button = gr.Button(label="Generate", value="Generate")
+        
+    with gr.Row():
+        gr.Examples(
+            examples=[
+                ['test_cases/hz0.png', 'test_cases/hz1.png'],
+                ['test_cases/more_cases/az0.png', 'test_cases/more_cases/az1.JPG'],
+                ['test_cases/more_cases/hi0.png', 'test_cases/more_cases/hi1.jpg'],
+                ['test_cases/more_cases/kn0.jpg', 'test_cases/more_cases/kn1.jpg'],
+                ['test_cases/more_cases/rk0.jpg', 'test_cases/more_cases/rk1.jpg'],
+            
+            
+            ],
+            inputs=[ref, base],
+            cache_examples=False,
+            examples_per_page=100
+        )
+        
+
+    run_local_button.click(fn=run_local, 
+                           inputs=[ref, 
+                                   base, 
+                                   ddim_steps, 
+                                   scale_ref,
+                                   scale_point, 
+                                   seed,
+                                   output_coords1,
+                                   output_coords2,
+                                   is_lineart
+                                  ], 
+                           outputs=[baseline_gallery]
+    )
+
+demo.launch(server_name="0.0.0.0")

scripts/infer.sh

@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+set -e
+set -x
+pretrained_model_name_or_path='./checkpoints/StableDiffusion'
+image_encoder_path='./checkpoints/models/clip-vit-large-patch14'
+controlnet_model_name_or_path='./checkpoints/models/control_v11p_sd15_lineart'
+annotator_ckpts_path='./checkpoints/models/Annotators'
+
+manga_reference_unet_path='./checkpoints/MangaNinjia/reference_unet.pth'
+manga_main_model_path='./checkpoints/MangaNinjia/denoising_unet.pth'
+manga_controlnet_model_path='./checkpoints/MangaNinjia/controlnet.pth'
+point_net_path='./checkpoints/MangaNinjia/point_net.pth'
+export CUDA_VISIBLE_DEVICES=0
+
+input_reference_paths='./test_cases/hz0.png ./test_cases/hz1.png'
+input_lineart_paths='./test_cases/hz1.png ./test_cases/hz0.png'
+point_ref_paths='./test_cases/hz01_0.npy ./test_cases/hz01_1.npy'
+point_lineart_paths='./test_cases/hz01_1.npy ./test_cases/hz01_0.npy'
+cd ..
+python infer.py  \
+    --seed 0 \
+    --denoise_steps 50 \
+    --pretrained_model_name_or_path $pretrained_model_name_or_path --image_encoder_path $image_encoder_path \
+    --controlnet_model_name_or_path $controlnet_model_name_or_path --annotator_ckpts_path $annotator_ckpts_path \
+    --manga_reference_unet_path $manga_reference_unet_path --manga_main_model_path $manga_main_model_path \
+    --manga_controlnet_model_path $manga_controlnet_model_path --point_net_path $point_net_path \
+    --output_dir 'output' \
+    --guidance_scale_ref 9 \
+    --guidance_scale_point 15 \
+    --input_reference_paths $input_reference_paths \
+    --input_lineart_paths $input_lineart_paths \
+    --point_ref_paths $point_ref_paths \
+    --point_lineart_paths $point_lineart_paths \

src/annotator/lineart/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Caroline Chan
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

src/annotator/lineart/__init__.py

@@ -0,0 +1,152 @@
+# From https://github.com/carolineec/informative-drawings
+# MIT License
+
+import os
+import cv2
+import torch
+import numpy as np
+
+import torch.nn as nn
+from einops import rearrange
+
+
+norm_layer = nn.InstanceNorm2d
+
+
+class ResidualBlock(nn.Module):
+    def __init__(self, in_features):
+        super(ResidualBlock, self).__init__()
+
+        conv_block = [  nn.ReflectionPad2d(1),
+                        nn.Conv2d(in_features, in_features, 3),
+                        norm_layer(in_features),
+                        nn.ReLU(inplace=True),
+                        nn.ReflectionPad2d(1),
+                        nn.Conv2d(in_features, in_features, 3),
+                        norm_layer(in_features)
+                        ]
+
+        self.conv_block = nn.Sequential(*conv_block)
+
+    def forward(self, x):
+        return x + self.conv_block(x)
+
+
+class Generator(nn.Module):
+    def __init__(self, input_nc, output_nc, n_residual_blocks=9, sigmoid=True):
+        super(Generator, self).__init__()
+
+        # Initial convolution block
+        model0 = [   nn.ReflectionPad2d(3),
+                    nn.Conv2d(input_nc, 64, 7),
+                    norm_layer(64),
+                    nn.ReLU(inplace=True) ]
+        self.model0 = nn.Sequential(*model0)
+
+        # Downsampling
+        model1 = []
+        in_features = 64
+        out_features = in_features*2
+        for _ in range(2):
+            model1 += [  nn.Conv2d(in_features, out_features, 3, stride=2, padding=1),
+                        norm_layer(out_features),
+                        nn.ReLU(inplace=True) ]
+            in_features = out_features
+            out_features = in_features*2
+        self.model1 = nn.Sequential(*model1)
+
+        model2 = []
+        # Residual blocks
+        for _ in range(n_residual_blocks):
+            model2 += [ResidualBlock(in_features)]
+        self.model2 = nn.Sequential(*model2)
+
+        # Upsampling
+        model3 = []
+        out_features = in_features//2
+        for _ in range(2):
+            model3 += [  nn.ConvTranspose2d(in_features, out_features, 3, stride=2, padding=1, output_padding=1),
+                        norm_layer(out_features),
+                        nn.ReLU(inplace=True) ]
+            in_features = out_features
+            out_features = in_features//2
+        self.model3 = nn.Sequential(*model3)
+
+        # Output layer
+        model4 = [  nn.ReflectionPad2d(3),
+                        nn.Conv2d(64, output_nc, 7)]
+        if sigmoid:
+            model4 += [nn.Sigmoid()]
+
+        self.model4 = nn.Sequential(*model4)
+
+    def forward(self, x, cond=None):
+        out = self.model0(x)
+        out = self.model1(out)
+        out = self.model2(out)
+        out = self.model3(out)
+        out = self.model4(out)
+
+        return out
+
+
+class LineartDetector:
+    def __init__(self, annotator_ckpts_path):
+        self.annotator_ckpts_path = annotator_ckpts_path
+        self.model = self.load_model('sk_model.pth')
+        self.model_coarse = self.load_model('sk_model2.pth')
+
+    def load_model(self, name):
+        remote_model_path = "https://huggingface.co/lllyasviel/Annotators/resolve/main/" + name
+        modelpath = os.path.join(self.annotator_ckpts_path, name)
+        if not os.path.exists(modelpath):
+            from basicsr.utils.download_util import load_file_from_url
+            load_file_from_url(remote_model_path, model_dir=self.annotator_ckpts_path)
+        model = Generator(3, 1, 3)
+        model.load_state_dict(torch.load(modelpath, map_location=torch.device('cpu')))
+        model.eval()
+        model = model.cuda()
+        return model
+
+    def __call__(self, input_image, coarse):
+        model = self.model_coarse if coarse else self.model
+        assert input_image.ndim == 3
+        image = input_image
+        with torch.no_grad():
+            image = torch.from_numpy(image).float().cuda()
+            image = image / 255.0
+            image = rearrange(image, 'h w c -> 1 c h w')
+            line = model(image)[0][0]
+
+            line = line.cpu().numpy()
+            line = (line * 255.0).clip(0, 255).astype(np.uint8)
+
+            return line
+
+class BatchLineartDetector:
+    def __init__(self, annotator_ckpts_path):
+        self.annotator_ckpts_path = annotator_ckpts_path
+        self.model = self.load_model('sk_model.pth')
+
+    def load_model(self, name):
+        remote_model_path = "https://huggingface.co/lllyasviel/Annotators/resolve/main/" + name
+        modelpath = os.path.join(self.annotator_ckpts_path, name)
+        if not os.path.exists(modelpath):
+            from basicsr.utils.download_util import load_file_from_url
+            load_file_from_url(remote_model_path, model_dir=self.annotator_ckpts_path)
+        model = Generator(3, 1, 3)
+        model.load_state_dict(torch.load(modelpath, map_location=torch.device('cpu')))
+        model.eval()
+        return model
+    
+    def to(self, device, dtype):
+        self.model.to(device, dtype=dtype)
+
+    def __call__(self, input_image, mean=-1., std=2.):
+        model = self.model
+        image = input_image
+        with torch.no_grad():
+            image = (image - mean) / std
+            line = model(image)
+        line = 1 - line
+        return line

src/models/attention.py

@@ -0,0 +1,444 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.models.attention import AdaLayerNorm, FeedForward
+from src.models.attention_processor import Attention
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from einops import rearrange
+from torch import nn
+
+
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(
+                dim, max_seq_length=num_positional_embeddings
+            )
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(
+                    dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+                )
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim
+                if not double_self_attention
+                else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(
+                dim, cross_attention_dim, num_attention_heads, attention_head_dim
+            )
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = (
+            cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        )
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states
+            if self.only_cross_attention
+            else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = (
+                norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+            )
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class TemporalBasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+
+        # SC-Attn
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = (
+            AdaLayerNorm(dim, num_embeds_ada_norm)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_ada_layer_norm_zero = False
+
+        # Temp-Attn
+        assert unet_use_temporal_attention is not None
+        if unet_use_temporal_attention:
+            self.attn_temp = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep)
+            if self.use_ada_layer_norm
+            else self.norm1(hidden_states)
+        )
+
+        if self.unet_use_cross_frame_attention:
+            hidden_states = (
+                self.attn1(
+                    norm_hidden_states,
+                    attention_mask=attention_mask,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+        else:
+            hidden_states = (
+                self.attn1(norm_hidden_states, attention_mask=attention_mask)
+                + hidden_states
+            )
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

src/models/mutual_self_attention_multi_scale.py

@@ -0,0 +1,389 @@
+# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/models/mutual_self_attention.py
+from typing import Any, Dict, Optional
+
+import torch
+from einops import rearrange
+
+from src.models.attention import TemporalBasicTransformerBlock
+
+from .attention import BasicTransformerBlock
+
+
+def torch_dfs(model: torch.nn.Module):
+    result = [model]
+    for child in model.children():
+        result += torch_dfs(child)
+    return result
+
+def filter_matrices_by_size(matrix_list, reference_matrix):
+    ref_shape = reference_matrix.shape[-2:]
+    return [matrix for matrix in matrix_list if matrix.shape[-2:] == ref_shape]
+
+class ReferenceAttentionControl:
+    def __init__(
+        self,
+        unet,
+        mode="write",
+        do_classifier_free_guidance=False,
+        attention_auto_machine_weight=float("inf"),
+        gn_auto_machine_weight=1.0,
+        style_fidelity=1.0,
+        reference_attn=True,
+        reference_adain=False,
+        fusion_blocks="midup",
+        batch_size=1,
+    ) -> None:
+        # 10. Modify self attention and group norm
+        self.unet = unet
+        assert mode in ["read", "write"]
+        assert fusion_blocks in ["midup", "full"]
+        self.reference_attn = reference_attn
+        self.reference_adain = reference_adain
+        self.fusion_blocks = fusion_blocks
+        self.register_reference_hooks(
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            fusion_blocks,
+            batch_size=batch_size,
+        )
+        self.point_embedding=[]
+    def register_reference_hooks(
+        self,
+        mode,
+        do_classifier_free_guidance,
+        attention_auto_machine_weight,
+        gn_auto_machine_weight,
+        style_fidelity,
+        reference_attn,
+        reference_adain,
+        dtype=torch.float16,
+        batch_size=1,
+        num_images_per_prompt=1,
+        device=torch.device("cpu"),
+        fusion_blocks="midup",
+    ):
+        MODE = mode
+        do_classifier_free_guidance = do_classifier_free_guidance
+        attention_auto_machine_weight = attention_auto_machine_weight
+        gn_auto_machine_weight = gn_auto_machine_weight
+        style_fidelity = style_fidelity
+        reference_attn = reference_attn
+        reference_adain = reference_adain
+        fusion_blocks = fusion_blocks
+        num_images_per_prompt = num_images_per_prompt
+        dtype = dtype
+        if do_classifier_free_guidance:
+            uc_mask = (
+                torch.Tensor(
+                    [1] * batch_size * num_images_per_prompt * 16
+                    + [0] * batch_size * num_images_per_prompt * 16
+                )
+                .to(device)
+                .bool()
+            )
+        else:
+            uc_mask = (
+                torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
+                .to(device)
+                .bool()
+            )
+
+        def hacked_basic_transformer_inner_forward(
+            self,
+            hidden_states: torch.FloatTensor,
+            attention_mask: Optional[torch.FloatTensor] = None,
+            encoder_hidden_states: Optional[torch.FloatTensor] = None,
+            encoder_attention_mask: Optional[torch.FloatTensor] = None,
+            timestep: Optional[torch.LongTensor] = None,
+            cross_attention_kwargs: Dict[str, Any] = None,
+            class_labels: Optional[torch.LongTensor] = None,
+            video_length=None,
+        ):
+            if self.use_ada_layer_norm:  # False
+                norm_hidden_states = self.norm1(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero:
+                (
+                    norm_hidden_states,
+                    gate_msa,
+                    shift_mlp,
+                    scale_mlp,
+                    gate_mlp,
+                ) = self.norm1(
+                    hidden_states,
+                    timestep,
+                    class_labels,
+                    hidden_dtype=hidden_states.dtype,
+                )
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+
+            # 1. Self-Attention
+            # self.only_cross_attention = False
+            cross_attention_kwargs = (
+                cross_attention_kwargs if cross_attention_kwargs is not None else {}
+            )
+            if self.only_cross_attention:
+                attn_output = self.attn1(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states
+                    if self.only_cross_attention
+                    else None,
+                    attention_mask=attention_mask,
+                    **cross_attention_kwargs,
+                )
+            else:
+                if MODE == "write":
+                    self.bank.append(norm_hidden_states.clone())
+                    attn_output = self.attn1(
+                        norm_hidden_states,
+                        encoder_hidden_states=encoder_hidden_states
+                        if self.only_cross_attention
+                        else None,
+                        attention_mask=attention_mask,
+                        **cross_attention_kwargs,
+                    )
+                if MODE == "read":
+                    bank_fea = [
+                        rearrange(
+                            d.unsqueeze(1).repeat(1, 1, 1, 1),
+                            "b t l c -> (b t) l c",
+                        )
+                        for d in self.bank
+                    ]
+                    try:
+                        modify_norm_hidden_states = torch.cat(
+                            [norm_hidden_states+self.point_bank_main[0].repeat(norm_hidden_states.shape[0],1,1)] + [bank_fea[0]+self.point_bank_ref[0].repeat(norm_hidden_states.shape[0],1,1)], dim=1
+                        )
+                        modify_norm_hidden_states_v = torch.cat(
+                            [norm_hidden_states] + bank_fea, dim=1
+                        )
+                        # import ipdb;ipdb.set_trace()
+                        hidden_states_uc = (
+                        self.attn1(
+                            norm_hidden_states+self.point_bank_main[0].repeat(norm_hidden_states.shape[0],1,1),
+                            encoder_hidden_states=modify_norm_hidden_states,
+                            encoder_hidden_states_v=modify_norm_hidden_states_v,
+                            attention_mask=attention_mask,
+                        )
+                        + hidden_states
+                        )
+                    except:
+                        modify_norm_hidden_states = torch.cat(
+                            [norm_hidden_states] + bank_fea, dim=1
+                        )
+                        hidden_states_uc = (
+                            self.attn1(
+                                norm_hidden_states,
+                                encoder_hidden_states=modify_norm_hidden_states,
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states
+                        )
+                    if do_classifier_free_guidance:
+                        hidden_states_c = hidden_states_uc.clone()
+                        _uc_mask = uc_mask.clone()
+                        if hidden_states.shape[0] != _uc_mask.shape[0]:
+                            _uc_mask = (
+                                torch.Tensor(
+                                    [1] * (hidden_states.shape[0] // 3)
+                                    + [0] * (hidden_states.shape[0] // 3)
+                                    + [0] * (hidden_states.shape[0] // 3)
+                                )
+                                .to(device)
+                                .bool()
+                            )
+                            _uc_mask_2 = (
+                                torch.Tensor(
+                                    [0] * (hidden_states.shape[0] // 3)
+                                    + [1] * (hidden_states.shape[0] // 3)
+                                    + [0] * (hidden_states.shape[0] // 3)
+                                )
+                                .to(device)
+                                .bool()
+                            )
+                        hidden_states_c[_uc_mask] = (
+                            self.attn1(
+                                norm_hidden_states[_uc_mask],
+                                encoder_hidden_states=norm_hidden_states[_uc_mask],
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states[_uc_mask]
+                        )
+                        modify_norm_hidden_states = torch.cat(
+                            [norm_hidden_states] + bank_fea, dim=1
+                        )
+                        hidden_states_c[_uc_mask_2] = (
+                            self.attn1(
+                                norm_hidden_states[_uc_mask_2],
+                                encoder_hidden_states=modify_norm_hidden_states[_uc_mask_2],
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states[_uc_mask_2]
+                        )
+                        hidden_states = hidden_states_c.clone()
+                    else:
+                        hidden_states = hidden_states_uc
+
+
+                    if self.attn2 is not None:
+                        # Cross-Attention
+                        norm_hidden_states = (
+                            self.norm2(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm2(hidden_states)
+                        )
+                        hidden_states = (
+                            self.attn2(
+                                norm_hidden_states,
+                                encoder_hidden_states=encoder_hidden_states,
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states
+                        )
+
+                    # Feed-forward
+                    hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+                    return hidden_states
+            # import ipdb;ipdb.set_trace()
+            if self.use_ada_layer_norm_zero:
+                attn_output = gate_msa.unsqueeze(1) * attn_output
+            try:
+                hidden_states = attn_output + hidden_states
+            except:
+                import ipdb;ipdb.set_trace()
+            if self.attn2 is not None:
+                norm_hidden_states = (
+                    self.norm2(hidden_states, timestep)
+                    if self.use_ada_layer_norm
+                    else self.norm2(hidden_states)
+                )
+
+                # 2. Cross-Attention
+                attn_output = self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=encoder_attention_mask,
+                    **cross_attention_kwargs,
+                )
+                hidden_states = attn_output + hidden_states
+
+            # 3. Feed-forward
+            norm_hidden_states = self.norm3(hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                norm_hidden_states = (
+                    norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+                )
+
+            ff_output = self.ff(norm_hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+            hidden_states = ff_output + hidden_states
+
+            return hidden_states
+
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            attn_modules = sorted(
+                attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+
+            for i, module in enumerate(attn_modules):
+                module._original_inner_forward = module.forward
+                if isinstance(module, BasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, BasicTransformerBlock
+                    )
+                module.bank = []
+                module.point_bank_ref=[]
+                module.point_bank_main=[]
+                module.attn_weight = float(i) / float(len(attn_modules))
+
+    def update(self, writer,point_embedding_ref=None,point_embedding_main=None,dtype=torch.float16):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(writer.unet.mid_block)
+                        + torch_dfs(writer.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in torch_dfs(writer.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            writer_attn_modules = sorted(
+                writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            # import ipdb;ipdb.set_trace()
+            for r, w in zip(reader_attn_modules, writer_attn_modules):
+                r.bank = [v.clone().to(dtype) for v in w.bank]
+                if point_embedding_main is not None:
+                    r.point_bank_ref=filter_matrices_by_size(point_embedding_ref, r.bank[0])
+                    r.point_bank_main=filter_matrices_by_size(point_embedding_main, r.bank[0])
+                # w.bank.clear()
+
+    def clear(self):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r in reader_attn_modules:
+                r.bank.clear()
+                r.point_bank_ref.clear()
+                r.point_bank_main.clear()

src/models/refunet_2d_condition.py

@@ -0,0 +1,1307 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+
+from .unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+    ref_features: Tuple[torch.FloatTensor] = None
+
+
+class RefUNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(
+            only_cross_attention
+        ) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if (
+            isinstance(transformer_layers_per_block, list)
+            and reverse_transformer_layers_per_block is None
+        ):
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError(
+                        "Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
+                    )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=conv_in_kernel,
+            padding=conv_in_padding,
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(
+                    f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
+                )
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2,
+                set_W_to_weight=False,
+                log=False,
+                flip_sin_to_cos=flip_sin_to_cos,
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(
+                block_out_channels[0], flip_sin_to_cos, freq_shift
+            )
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info(
+                "encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
+            )
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(
+                timestep_input_dim, time_embed_dim, act_fn=act_fn
+            )
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim,
+                time_embed_dim,
+                num_heads=addition_embed_type_num_heads,
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim,
+                image_embed_dim=cross_attention_dim,
+                time_embed_dim=time_embed_dim,
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(
+                addition_time_embed_dim, flip_sin_to_cos, freq_shift
+            )
+            self.add_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type is not None:
+            raise ValueError(
+                f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
+            )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(
+                down_block_types
+            )
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i]
+                if attention_head_dim[i] is not None
+                else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+        self.conv_norm_out = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        # self.conv_out = nn.Conv2d(
+        #     block_out_channels[0],
+        #     out_channels,
+        #     kernel_size=conv_out_kernel,
+        #     padding=conv_out_padding,
+        # )
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(
+                cross_attention_dim, list
+            ):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            # self.position_net = PositionNet(
+            #     positive_len=positive_len,
+            #     out_dim=cross_attention_dim,
+            #     feature_type=feature_type,
+            # )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(
+                    return_deprecated_lora=True
+                )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+        self,
+        processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
+        _remove_lora=False,
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(
+                        processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
+                    )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(
+            proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnAddedKVProcessor()
+        elif all(
+            proc.__class__ in CROSS_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = (
+            num_sliceable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+            module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if (
+                    hasattr(upsample_block, k)
+                    or getattr(upsample_block, k, None) is not None
+                ):
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(sample.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if (
+                "image_embeds" not in added_cond_kwargs
+                or "hint" not in added_cond_kwargs
+            ):
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_proj"
+        ):
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_image_proj"
+        ):
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(
+                encoder_hidden_states, image_embeds
+            )
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "image_proj"
+        ):
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "ip_image_proj"
+        ):
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds).to(
+                encoder_hidden_states.dtype
+            )
+            encoder_hidden_states = torch.cat(
+                [encoder_hidden_states, image_embeds], dim=1
+            )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # # 2.5 GLIGEN position net
+        # if (
+        #     cross_attention_kwargs is not None
+        #     and cross_attention_kwargs.get("gligen", None) is not None
+        # ):
+        #     cross_attention_kwargs = cross_attention_kwargs.copy()
+        #     gligen_args = cross_attention_kwargs.pop("gligen")
+        #     cross_attention_kwargs["gligen"] = {
+        #         "objs": self.position_net(**gligen_args)
+        #     }
+
+        # 3. down
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = (
+            mid_block_additional_residual is not None
+            and down_block_additional_residuals is not None
+        )
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if (
+            not is_adapter
+            and mid_block_additional_residual is None
+            and down_block_additional_residuals is not None
+        ):
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        tot_referece_features = ()
+        for downsample_block in self.down_blocks:
+            if (
+                hasattr(downsample_block, "has_cross_attention")
+                and downsample_block.has_cross_attention
+            ):
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals[
+                        "additional_residuals"
+                    ] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample, temb=emb, scale=lora_scale
+                )
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = (
+                    down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples = new_down_block_res_samples + (
+                    down_block_res_sample,
+                )
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if (
+                hasattr(self.mid_block, "has_cross_attention")
+                and self.mid_block.has_cross_attention
+            ):
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[
+                : -len(upsample_block.resnets)
+            ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                hasattr(upsample_block, "has_cross_attention")
+                and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+
+        # 6. post-process
+        # if self.conv_norm_out:
+        #     sample = self.conv_norm_out(sample)
+        #     sample = self.conv_act(sample)
+        # sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

src/models/transformer_2d.py

@@ -0,0 +1,396 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformer_2d.py
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+# from diffusers.models.embeddings import CaptionProjection
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from torch import nn
+
+from .attention import BasicTransformerBlock
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+    ref_feature: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate(
+                "norm_type!=num_embeds_ada_norm",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif (
+            not self.is_input_continuous
+            and not self.is_input_vectorized
+            and not self.is_input_patches
+        ):
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups,
+            num_channels=in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        if use_linear_projection:
+            self.proj_in = linear_cls(in_channels, inner_dim)
+        else:
+            self.proj_in = conv_cls(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        # TODO: should use out_channels for continuous projections
+        if use_linear_projection:
+            self.proj_out = linear_cls(inner_dim, in_channels)
+        else:
+            self.proj_out = conv_cls(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(
+                inner_dim, use_additional_conditions=self.use_additional_conditions
+            )
+
+        self.caption_projection = None
+        # if caption_channels is not None:
+        #     self.caption_projection = CaptionProjection(
+        #         in_features=caption_channels, hidden_size=inner_dim
+        #     )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(hidden_states.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 1. Input
+        batch, _, height, width = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(
+                batch_size, -1, hidden_states.shape[-1]
+            )
+
+        ref_feature = hidden_states.reshape(batch, height, width, inner_dim)
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+
+            output = hidden_states + residual
+        if not return_dict:
+            return (output, ref_feature)
+
+        return Transformer2DModelOutput(sample=output, ref_feature=ref_feature)

src/models/unet_2d_blocks.py

@@ -0,0 +1,1131 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+from typing import Any, Dict, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import Attention
+from diffusers.models.dual_transformer_2d import DualTransformer2DModel
+from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import apply_freeu
+from torch import nn
+
+from .transformer_2d import Transformer2DModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_down_block(
+    down_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    temb_channels: int,
+    add_downsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    downsample_padding: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    downsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock2D"
+            )
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    prev_output_channel: int,
+    temb_channels: int,
+    add_upsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resolution_idx: Optional[int] = None,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    upsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+) -> nn.Module:
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock2D"
+            )
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+
+    raise ValueError(f"{up_block_type} does not exist.")
+def get_mid_block(
+    mid_block_type: str,
+    temb_channels: int,
+    in_channels: int,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resnet_groups: int,
+    output_scale_factor: float = 1.0,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    mid_block_only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = 1,
+    dropout: float = 0.0,
+):
+    if mid_block_type == "UNetMidBlock2DCrossAttn":
+        return UNetMidBlock2DCrossAttn(
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            output_scale_factor=output_scale_factor,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            resnet_groups=resnet_groups,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            upcast_attention=upcast_attention,
+            attention_type=attention_type,
+        )
+    elif mid_block_type == "UNetMidBlock2D":
+        return UNetMidBlock2D(
+            in_channels=in_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            num_layers=0,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            output_scale_factor=output_scale_factor,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            add_attention=False,
+        )
+    elif mid_block_type is None:
+        return None
+    else:
+        raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+
+class AutoencoderTinyBlock(nn.Module):
+    """
+    Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
+    blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        out_channels (`int`): The number of output channels.
+        act_fn (`str`):
+            ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
+
+    Returns:
+        `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
+        `out_channels`.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+        super().__init__()
+        act_fn = get_activation(act_fn)
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+        )
+        self.skip = (
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+        self.fuse = nn.ReLU()
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+    """
+    A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        temb_channels (`int`): The number of temporal embedding channels.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+        num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+        resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+        resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+            The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+            model on tasks with long-range temporal dependencies.
+        resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+        resnet_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use in the group normalization layers of the resnet blocks.
+        attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+        resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+            Whether to use pre-normalization for the resnet blocks.
+        add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+        attention_head_dim (`int`, *optional*, defaults to 1):
+            Dimension of a single attention head. The number of attention heads is determined based on this value and
+            the number of input channels.
+        output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+    Returns:
+        `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+        in_channels, height, width)`.
+
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        attn_groups: Optional[int] = None,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+        self.add_attention = add_attention
+
+        if attn_groups is None:
+            attn_groups = (
+                resnet_groups if resnet_time_scale_shift == "default" else None
+            )
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warn(
+                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=attn_groups,
+                        spatial_norm_dim=temb_channels
+                        if resnet_time_scale_shift == "spatial"
+                        else None,
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None
+    ) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = attn(hidden_states, temb=temb)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        output_scale_factor: float = 1.0,
+        cross_attention_dim: int = 1280,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        # support for variable transformer layers per block
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for i in range(num_layers):
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        downsample_padding: int = 1,
+        add_downsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        additional_residuals: Optional[torch.FloatTensor] = None,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        blocks = list(zip(self.resnets, self.attentions))
+
+        for i, (resnet, attn) in enumerate(blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+            # apply additional residuals to the output of the last pair of resnet and attention blocks
+            if i == len(blocks) - 1 and additional_residuals is not None:
+                hidden_states = hidden_states + additional_residuals
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_padding: int = 1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        upsample_size: Optional[int] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(
+                    hidden_states, upsample_size, scale=lora_scale
+                )
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        upsample_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+        return hidden_states

src/models/unet_2d_condition.py

@@ -0,0 +1,1305 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    # PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+
+from .unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+    get_up_block,
+    get_mid_block
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
+        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
+        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: float = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads: int = 64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        self._check_config(
+            down_block_types=down_block_types,
+            up_block_types=up_block_types,
+            only_cross_attention=only_cross_attention,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            cross_attention_dim=cross_attention_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+        )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        time_embed_dim, timestep_input_dim = self._set_time_proj(
+            time_embedding_type,
+            block_out_channels=block_out_channels,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            time_embedding_dim=time_embedding_dim,
+        )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        self._set_encoder_hid_proj(
+            encoder_hid_dim_type,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+        )
+
+        # class embedding
+        self._set_class_embedding(
+            class_embed_type,
+            act_fn=act_fn,
+            num_class_embeds=num_class_embeds,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+            timestep_input_dim=timestep_input_dim,
+        )
+
+        self._set_add_embedding(
+            addition_embed_type,
+            addition_embed_type_num_heads=addition_embed_type_num_heads,
+            addition_time_embed_dim=addition_time_embed_dim,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+        )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = get_mid_block(
+            mid_block_type,
+            temb_channels=blocks_time_embed_dim,
+            in_channels=block_out_channels[-1],
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            output_scale_factor=mid_block_scale_factor,
+            transformer_layers_per_block=transformer_layers_per_block[-1],
+            num_attention_heads=num_attention_heads[-1],
+            cross_attention_dim=cross_attention_dim[-1],
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            mid_block_only_cross_attention=mid_block_only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+            resnet_skip_time_act=resnet_skip_time_act,
+            cross_attention_norm=cross_attention_norm,
+            attention_head_dim=attention_head_dim[-1],
+            dropout=dropout,
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+        )
+
+        self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)
+
+    def _check_config(
+        self,
+        down_block_types: Tuple[str],
+        up_block_types: Tuple[str],
+        only_cross_attention: Union[bool, Tuple[bool]],
+        block_out_channels: Tuple[int],
+        layers_per_block: Union[int, Tuple[int]],
+        cross_attention_dim: Union[int, Tuple[int]],
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
+        reverse_transformer_layers_per_block: bool,
+        attention_head_dim: int,
+        num_attention_heads: Optional[Union[int, Tuple[int]]],
+    ):
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+    def _set_time_proj(
+        self,
+        time_embedding_type: str,
+        block_out_channels: int,
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        time_embedding_dim: int,
+    ) -> Tuple[int, int]:
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        return time_embed_dim, timestep_input_dim
+
+    def _set_encoder_hid_proj(
+        self,
+        encoder_hid_dim_type: Optional[str],
+        cross_attention_dim: Union[int, Tuple[int]],
+        encoder_hid_dim: Optional[int],
+    ):
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+    def _set_class_embedding(
+        self,
+        class_embed_type: Optional[str],
+        act_fn: str,
+        num_class_embeds: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+        timestep_input_dim: int,
+    ):
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+    def _set_add_embedding(
+        self,
+        addition_embed_type: str,
+        addition_embed_type_num_heads: int,
+        addition_time_embed_dim: Optional[int],
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        cross_attention_dim: Optional[int],
+        encoder_hid_dim: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+    ):
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+    def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = GLIGENTextBoundingboxProjection(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def unload_lora(self):
+        """Unloads LoRA weights."""
+        deprecate(
+            "unload_lora",
+            "0.28.0",
+            "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().",
+        )
+        for module in self.modules():
+            if hasattr(module, "set_lora_layer"):
+                module.set_lora_layer(None)
+
+    def get_time_embed(
+        self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
+    ) -> Optional[torch.Tensor]:
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+        return t_emb
+
+    def get_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
+        class_emb = None
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+        return class_emb
+
+    def get_aug_embed(
+        self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> Optional[torch.Tensor]:
+        aug_emb = None
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb = self.add_embedding(image_embs, hint)
+        return aug_emb
+
+    def process_encoder_hidden_states(
+        self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> torch.Tensor:
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds)
+            encoder_hidden_states = (encoder_hidden_states, image_embeds)
+        return encoder_hidden_states
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        t_emb = self.get_time_embed(sample=sample, timestep=timestep)
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+        class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
+        if class_emb is not None:
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        aug_emb = self.get_aug_embed(
+            emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+        if self.config.addition_embed_type == "image_hint":
+            aug_emb, hint = aug_emb
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        encoder_hidden_states = self.process_encoder_hidden_states(
+            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+        # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
+        # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
+        if cross_attention_kwargs is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            lora_scale = cross_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

src/point_network.py

@@ -0,0 +1,45 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Tuple
+from diffusers.models.modeling_utils import ModelMixin
+class PointNet(ModelMixin):
+    def __init__(
+        self,
+        conditioning_channels: int = 1,
+        out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        downsamples: Tuple[int] = (6, 2, 2, 2)
+    ):
+        super(PointNet, self).__init__()
+        
+        self.blocks = nn.ModuleList()
+        current_channels = conditioning_channels
+        
+        # 构造卷积块
+        for out_channel, downsample in zip(out_channels, downsamples):
+            layers = []
+            for _ in range(downsample // 2):
+                layers.append(nn.Conv2d(in_channels=current_channels, out_channels=out_channel, kernel_size=3, stride=2, padding=1))
+                layers.append(nn.SiLU())
+                current_channels = out_channel
+            self.blocks.append(nn.Sequential(*layers))
+    
+    def forward(self, x):
+        embeddings = []
+        embedding = x
+        for block in self.blocks:
+            embedding = block(embedding)
+            B, C, H, W = embedding.shape 
+            embeddings.append(embedding.view(B, C, H * W).transpose(1, 2))
+            # embeddings.append(embedding)
+        return embeddings
+
+if __name__ == "__main__":
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    print(f'Using device: {device}')
+    model = PointNet().to(device)
+    
+    dummy_input = torch.randn(1, 1, 288, 512).to(device)  # Batch size = 1, Channels = 1, Height = 288, Width = 512
+    embeddings = model(dummy_input)
+    for i, embedding in enumerate(embeddings):
+        print(f"Output at layer {i + 1}:", embedding.shape)

test_cases/hz01_0.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:810f8089bdc9833596ef163a372128de1b5e9e7fce29e17779e2d0539119d10e
+size 262272

test_cases/hz01_1.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72334ed6a7ab06e43b554cb3ce98830e012270762aa64fecfc1e7eaa49037019
+size 262272

utils/image_util.py

@@ -0,0 +1,36 @@
+import matplotlib
+import numpy as np
+import torch
+from PIL import Image
+
+def resize_max_res(img: Image.Image, max_edge_resolution: int) -> Image.Image:
+    """
+    Resize image to limit maximum edge length while keeping aspect ratio.
+    Args:
+        img (`Image.Image`):
+            Image to be resized.
+        max_edge_resolution (`int`):
+            Maximum edge length (pixel).
+    Returns:
+        `Image.Image`: Resized image.
+    """
+    
+    original_width, original_height = img.size
+    
+    downscale_factor = min(
+        max_edge_resolution / original_width, max_edge_resolution / original_height
+    )
+
+    new_width = int(original_width * downscale_factor)
+    new_height = int(original_height * downscale_factor)
+
+    resized_img = img.resize((new_width, new_height))
+    return resized_img
+
+def chw2hwc(chw):
+    assert 3 == len(chw.shape)
+    if isinstance(chw, torch.Tensor):
+        hwc = torch.permute(chw, (1, 2, 0))
+    elif isinstance(chw, np.ndarray):
+        hwc = np.moveaxis(chw, 0, -1)
+    return hwc