---
tags:
- text-to-image
- pytorch
---
# DDT: Decoupled Diffusion Transformer
# DDT: Decoupled Diffusion Transformer
## Introduction
We decouple diffusion transformer into encoder-decoder design, and surpresingly that a **more substantial encoder yields performance improvements as model size increases**.
* We achieves **1.26 FID** on ImageNet256x256 Benchmark with DDT-XL/2(22en6de).
* We achieves **1.28 FID** on ImageNet512x512 Benchmark with DDT-XL/2(22en6de).
* As a byproduct, our DDT can reuse encoder among adjacent steps to accelerate inference.
## Visualizations
## Checkpoints
We take the off-shelf [VAE](https://huggingface.co/stabilityai/sd-vae-ft-ema) to encode image into latent space, and train the decoder with DDT.
| Dataset | Model | Params | FID | HuggingFace |
|-------------|-------------------|-----------|------|----------------------------------------------------------|
| ImageNet256 | DDT-XL/2(22en6de) | 675M | 1.26 | [🤗](https://huggingface.co/MCG-NJU/DDT-XL-22en6de-R256) |
| ImageNet512 | DDT-XL/2(22en6de) | 675M | 1.28 | [🤗](https://huggingface.co/MCG-NJU/DDT-XL-22en6de-R512) |
## Online Demos
Coming soon.
## Usages
We use ADM evaluation suite to report FID.
```bash
# for installation
pip install -r requirements.txt
```
```bash
# for inference
python main.py predict -c configs/repa_improved_ddt_xlen22de6_256.yaml --ckpt_path=XXX.ckpt
```
```bash
# for training
# extract image latent (optional)
python3 tools/cache_imlatent4.py
# train
python main.py fit -c configs/repa_improved_ddt_xlen22de6_256.yaml
```
## Reference
```bibtex
@ARTICLE{ddt,
title = "DDT: Decoupled Diffusion Transformer",
author = "Wang, Shuai and Tian, Zhi and Huang, Weilin and Wang, Limin",
month = apr,
year = 2025,
archivePrefix = "arXiv",
primaryClass = "cs.CV",
eprint = "2504.05741"
}
```
