Upload 7 files

flux/block.py

@@ -0,0 +1,461 @@
+# Recycled from Ominicontrol and modified to accept an extra condition.
+# While Zenctrl pursued a similar idea, it diverged structurally. 
+# We appreciate the clarity of Omini's implementation and decided to align with it.
+
+import torch
+from typing import List, Union, Optional, Dict, Any, Callable
+from diffusers.models.attention_processor import Attention, F
+from .lora_controller import enable_lora
+from diffusers.models.embeddings import apply_rotary_emb
+
+def attn_forward(
+    attn: Attention,
+    hidden_states: torch.FloatTensor,
+    encoder_hidden_states: torch.FloatTensor = None,
+    condition_latents: torch.FloatTensor = None,
+    extra_condition_latents: torch.FloatTensor = None,
+    attention_mask: Optional[torch.FloatTensor] = None,
+    image_rotary_emb: Optional[torch.Tensor] = None,
+    cond_rotary_emb: Optional[torch.Tensor] = None,
+    extra_cond_rotary_emb: Optional[torch.Tensor] = None,
+    model_config: Optional[Dict[str, Any]] = {},
+) -> torch.FloatTensor:
+    batch_size, _, _ = (
+        hidden_states.shape
+        if encoder_hidden_states is None
+        else encoder_hidden_states.shape
+    )
+
+    with enable_lora(
+        (attn.to_q, attn.to_k, attn.to_v), model_config.get("latent_lora", False)
+    ):
+        # `sample` projections.
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+    inner_dim = key.shape[-1]
+    head_dim = inner_dim // attn.heads
+
+    query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+    key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+    value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+    if attn.norm_q is not None:
+        query = attn.norm_q(query)
+    if attn.norm_k is not None:
+        key = attn.norm_k(key)
+
+    # the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
+    if encoder_hidden_states is not None:
+        # `context` projections.
+        encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+        encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
+            batch_size, -1, attn.heads, head_dim
+        ).transpose(1, 2)
+        encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
+            batch_size, -1, attn.heads, head_dim
+        ).transpose(1, 2)
+        encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
+            batch_size, -1, attn.heads, head_dim
+        ).transpose(1, 2)
+
+        if attn.norm_added_q is not None:
+            encoder_hidden_states_query_proj = attn.norm_added_q(
+                encoder_hidden_states_query_proj
+            )
+        if attn.norm_added_k is not None:
+            encoder_hidden_states_key_proj = attn.norm_added_k(
+                encoder_hidden_states_key_proj
+            )
+
+        # attention
+        query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
+        key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+        value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+
+    if image_rotary_emb is not None:
+        
+
+        query = apply_rotary_emb(query, image_rotary_emb)
+        key = apply_rotary_emb(key, image_rotary_emb)
+
+    if condition_latents is not None:
+        cond_query = attn.to_q(condition_latents)
+        cond_key = attn.to_k(condition_latents)
+        cond_value = attn.to_v(condition_latents)
+
+        cond_query = cond_query.view(batch_size, -1, attn.heads, head_dim).transpose(
+            1, 2
+        )
+        cond_key = cond_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        cond_value = cond_value.view(batch_size, -1, attn.heads, head_dim).transpose(
+            1, 2
+        )
+        if attn.norm_q is not None:
+            cond_query = attn.norm_q(cond_query)
+        if attn.norm_k is not None:
+            cond_key = attn.norm_k(cond_key)
+    
+    #extra condition
+    if extra_condition_latents is not None:
+        extra_cond_query = attn.to_q(extra_condition_latents)
+        extra_cond_key = attn.to_k(extra_condition_latents)
+        extra_cond_value = attn.to_v(extra_condition_latents)
+
+        extra_cond_query = extra_cond_query.view(batch_size, -1, attn.heads, head_dim).transpose(
+            1, 2
+        )
+        extra_cond_key = extra_cond_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        extra_cond_value = extra_cond_value.view(batch_size, -1, attn.heads, head_dim).transpose(
+            1, 2
+        )
+        if attn.norm_q is not None:
+            extra_cond_query = attn.norm_q(extra_cond_query)
+        if attn.norm_k is not None:
+            extra_cond_key = attn.norm_k(extra_cond_key)
+
+
+    if extra_cond_rotary_emb is not None:
+        extra_cond_query = apply_rotary_emb(extra_cond_query, extra_cond_rotary_emb)
+        extra_cond_key = apply_rotary_emb(extra_cond_key, extra_cond_rotary_emb)
+
+    if cond_rotary_emb is not None:
+        cond_query = apply_rotary_emb(cond_query, cond_rotary_emb)
+        cond_key = apply_rotary_emb(cond_key, cond_rotary_emb)
+
+    if condition_latents is not None:
+        if extra_condition_latents is not None:
+            
+            query = torch.cat([query, cond_query, extra_cond_query], dim=2)
+            key = torch.cat([key, cond_key, extra_cond_key], dim=2)
+            value = torch.cat([value, cond_value, extra_cond_value], dim=2)
+        else:
+            query = torch.cat([query, cond_query], dim=2)
+            key = torch.cat([key, cond_key], dim=2)
+            value = torch.cat([value, cond_value], dim=2)
+            print("concat Omini latents: ", query.shape, key.shape, value.shape)
+
+        
+    if not model_config.get("union_cond_attn", True):
+        
+        attention_mask = torch.ones(
+            query.shape[2], key.shape[2], device=query.device, dtype=torch.bool
+        )
+        condition_n = cond_query.shape[2]
+        attention_mask[-condition_n:, :-condition_n] = False
+        attention_mask[:-condition_n, -condition_n:] = False
+    elif model_config.get("independent_condition", False):
+        attention_mask = torch.ones(
+            query.shape[2], key.shape[2], device=query.device, dtype=torch.bool
+        )
+        condition_n = cond_query.shape[2]
+        attention_mask[-condition_n:, :-condition_n] = False
+        
+    if hasattr(attn, "c_factor"):
+        attention_mask = torch.zeros(
+            query.shape[2], key.shape[2], device=query.device, dtype=query.dtype
+        )
+        condition_n = cond_query.shape[2]
+        condition_e = extra_cond_query.shape[2]
+        bias = torch.log(attn.c_factor[0])
+        attention_mask[-condition_n-condition_e:-condition_e, :-condition_n-condition_e] = bias
+        attention_mask[:-condition_n-condition_e, -condition_n-condition_e:-condition_e] = bias
+        
+        bias = torch.log(attn.c_factor[1])
+        attention_mask[-condition_e:, :-condition_n-condition_e] = bias
+        attention_mask[:-condition_n-condition_e, -condition_e:] = bias
+        
+    hidden_states = F.scaled_dot_product_attention(
+        query, key, value, dropout_p=0.0, is_causal=False, attn_mask=attention_mask
+    )
+    hidden_states = hidden_states.transpose(1, 2).reshape(
+        batch_size, -1, attn.heads * head_dim
+    )
+    hidden_states = hidden_states.to(query.dtype)
+
+    if encoder_hidden_states is not None:
+        if condition_latents is not None:
+            if extra_condition_latents is not None:
+                encoder_hidden_states, hidden_states, condition_latents, extra_condition_latents = (
+                    hidden_states[:, : encoder_hidden_states.shape[1]],
+                    hidden_states[
+                        :, encoder_hidden_states.shape[1] : -condition_latents.shape[1]*2
+                    ],
+                    hidden_states[:, -condition_latents.shape[1]*2 :-condition_latents.shape[1]],
+                    hidden_states[:, -condition_latents.shape[1] :], #extra condition latents
+                )
+            else:
+                encoder_hidden_states, hidden_states, condition_latents = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[
+                    :, encoder_hidden_states.shape[1] : -condition_latents.shape[1]
+                ],
+                hidden_states[:, -condition_latents.shape[1] :]
+            )
+        else:
+            encoder_hidden_states, hidden_states = (
+                hidden_states[:, : encoder_hidden_states.shape[1]],
+                hidden_states[:, encoder_hidden_states.shape[1] :],
+            )
+
+        with enable_lora((attn.to_out[0],), model_config.get("latent_lora", False)):
+            # linear proj
+            hidden_states = attn.to_out[0](hidden_states)
+            # dropout
+            hidden_states = attn.to_out[1](hidden_states)
+        encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        if condition_latents is not None:
+            condition_latents = attn.to_out[0](condition_latents)
+            condition_latents = attn.to_out[1](condition_latents)
+
+        if extra_condition_latents is not None:
+            extra_condition_latents = attn.to_out[0](extra_condition_latents)
+            extra_condition_latents = attn.to_out[1](extra_condition_latents)
+
+
+        return (
+            # (hidden_states, encoder_hidden_states, condition_latents, extra_condition_latents)
+            (hidden_states, encoder_hidden_states, condition_latents, extra_condition_latents)
+            if condition_latents is not None
+            else (hidden_states, encoder_hidden_states)
+        )
+    elif condition_latents is not None:
+        # if there are condition_latents, we need to separate the hidden_states and the condition_latents
+        if extra_condition_latents is not None:
+            hidden_states, condition_latents, extra_condition_latents = (
+                hidden_states[:, : -condition_latents.shape[1]*2],
+                hidden_states[:, -condition_latents.shape[1]*2 :-condition_latents.shape[1]],
+                hidden_states[:, -condition_latents.shape[1] :],
+            )
+        else:
+            hidden_states, condition_latents = (
+                hidden_states[:, : -condition_latents.shape[1]],
+                hidden_states[:, -condition_latents.shape[1] :],
+            )
+        return hidden_states, condition_latents, extra_condition_latents
+    else:
+        return hidden_states
+
+
+def block_forward(
+    self,
+    hidden_states: torch.FloatTensor,
+    encoder_hidden_states: torch.FloatTensor,
+    condition_latents: torch.FloatTensor,
+    extra_condition_latents: torch.FloatTensor,
+    temb: torch.FloatTensor,
+    cond_temb: torch.FloatTensor,
+    extra_cond_temb: torch.FloatTensor,
+    cond_rotary_emb=None,
+    extra_cond_rotary_emb=None,
+    image_rotary_emb=None,
+    model_config: Optional[Dict[str, Any]] = {},
+):
+    use_cond = condition_latents is not None
+
+    use_extra_cond = extra_condition_latents is not None
+    with enable_lora((self.norm1.linear,), model_config.get("latent_lora", False)):
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+            hidden_states, emb=temb
+        )
+
+    norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = (
+        self.norm1_context(encoder_hidden_states, emb=temb)
+    )
+
+    if use_cond:
+        (
+            norm_condition_latents,
+            cond_gate_msa,
+            cond_shift_mlp,
+            cond_scale_mlp,
+            cond_gate_mlp,
+        ) = self.norm1(condition_latents, emb=cond_temb)
+        (
+            norm_extra_condition_latents,
+            extra_cond_gate_msa,
+            extra_cond_shift_mlp,
+            extra_cond_scale_mlp,
+            extra_cond_gate_mlp,
+        ) = self.norm1(extra_condition_latents, emb=extra_cond_temb)
+
+    # Attention.
+    result = attn_forward(
+        self.attn,
+        model_config=model_config,
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        condition_latents=norm_condition_latents if use_cond else None,
+        extra_condition_latents=norm_extra_condition_latents if use_cond else None,
+        image_rotary_emb=image_rotary_emb,
+        cond_rotary_emb=cond_rotary_emb if use_cond else None,
+        extra_cond_rotary_emb=extra_cond_rotary_emb if use_extra_cond else None,
+    )
+    # print("in self block: ", result.shape)
+    attn_output, context_attn_output = result[:2]
+    cond_attn_output = result[2] if use_cond else None
+    extra_condition_output = result[3]
+
+    # Process attention outputs for the `hidden_states`.
+    # 1. hidden_states
+    attn_output = gate_msa.unsqueeze(1) * attn_output
+    hidden_states = hidden_states + attn_output
+    # 2. encoder_hidden_states
+    context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
+    
+    encoder_hidden_states = encoder_hidden_states + context_attn_output 
+    # 3. condition_latents
+    if use_cond:
+        cond_attn_output = cond_gate_msa.unsqueeze(1) * cond_attn_output
+        condition_latents = condition_latents + cond_attn_output
+        #need to make new condition_extra and add extra_condition_output
+        if use_extra_cond:
+            extra_condition_output = extra_cond_gate_msa.unsqueeze(1) * extra_condition_output
+            extra_condition_latents = extra_condition_latents + extra_condition_output
+
+        if model_config.get("add_cond_attn", False):
+            hidden_states += cond_attn_output
+            hidden_states += extra_condition_output
+
+
+    # LayerNorm + MLP.
+    # 1. hidden_states
+    norm_hidden_states = self.norm2(hidden_states)
+    norm_hidden_states = (
+        norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+    )
+    # 2. encoder_hidden_states
+    norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+    norm_encoder_hidden_states = (
+        norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+    )
+    # 3. condition_latents
+    if use_cond:
+        norm_condition_latents = self.norm2(condition_latents)
+        norm_condition_latents = (
+            norm_condition_latents * (1 + cond_scale_mlp[:, None])
+            + cond_shift_mlp[:, None]
+        )
+
+    if use_extra_cond:
+        #added conditions
+        extra_norm_condition_latents = self.norm2(extra_condition_latents)
+        extra_norm_condition_latents = (
+            extra_norm_condition_latents * (1 + extra_cond_scale_mlp[:, None])
+            + extra_cond_shift_mlp[:, None]
+        )
+    
+    # Feed-forward.
+    with enable_lora((self.ff.net[2],), model_config.get("latent_lora", False)):
+        # 1. hidden_states
+        ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+    # 2. encoder_hidden_states
+    context_ff_output = self.ff_context(norm_encoder_hidden_states)
+    context_ff_output = c_gate_mlp.unsqueeze(1) * context_ff_output
+    # 3. condition_latents
+    if use_cond:
+        cond_ff_output = self.ff(norm_condition_latents)
+        cond_ff_output = cond_gate_mlp.unsqueeze(1) * cond_ff_output
+
+    if use_extra_cond:
+        extra_cond_ff_output = self.ff(extra_norm_condition_latents)
+        extra_cond_ff_output = extra_cond_gate_mlp.unsqueeze(1) * extra_cond_ff_output
+
+    # Process feed-forward outputs.
+    hidden_states = hidden_states + ff_output
+    encoder_hidden_states = encoder_hidden_states + context_ff_output
+    if use_cond:
+        condition_latents = condition_latents + cond_ff_output
+    if use_extra_cond:
+        extra_condition_latents = extra_condition_latents + extra_cond_ff_output
+
+    # Clip to avoid overflow.
+    if encoder_hidden_states.dtype == torch.float16:
+        encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+
+    return encoder_hidden_states, hidden_states, condition_latents, extra_condition_latents if use_cond else None
+
+
+def single_block_forward(
+    self,
+    hidden_states: torch.FloatTensor,
+    temb: torch.FloatTensor,
+    image_rotary_emb=None,
+    condition_latents: torch.FloatTensor = None,
+    extra_condition_latents: torch.FloatTensor = None,
+    cond_temb: torch.FloatTensor = None,
+    extra_cond_temb: torch.FloatTensor = None,
+    cond_rotary_emb=None,
+    extra_cond_rotary_emb=None,
+    model_config: Optional[Dict[str, Any]] = {},
+):
+
+    using_cond = condition_latents is not None
+    using_extra_cond = extra_condition_latents is not None
+    residual = hidden_states
+    with enable_lora(
+        (
+            self.norm.linear,
+            self.proj_mlp,
+        ),
+        model_config.get("latent_lora", False),
+    ):
+        norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
+        mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
+    if using_cond:
+        residual_cond = condition_latents
+        norm_condition_latents, cond_gate = self.norm(condition_latents, emb=cond_temb)
+        mlp_cond_hidden_states = self.act_mlp(self.proj_mlp(norm_condition_latents))
+
+    if using_extra_cond:
+        extra_residual_cond = extra_condition_latents
+        extra_norm_condition_latents, extra_cond_gate = self.norm(extra_condition_latents, emb=extra_cond_temb)
+        extra_mlp_cond_hidden_states = self.act_mlp(self.proj_mlp(extra_norm_condition_latents))
+
+    attn_output = attn_forward(
+        self.attn,
+        model_config=model_config,
+        hidden_states=norm_hidden_states,
+        image_rotary_emb=image_rotary_emb,
+        **(
+            {
+                "condition_latents": norm_condition_latents,
+                "cond_rotary_emb": cond_rotary_emb if using_cond else None,
+                "extra_condition_latents": extra_norm_condition_latents if using_cond else None,
+                "extra_cond_rotary_emb": extra_cond_rotary_emb if using_cond else None,
+            }
+            if using_cond
+            else {}
+        ),
+    )
+
+    if using_cond:
+        attn_output, cond_attn_output, extra_cond_attn_output = attn_output
+
+
+    with enable_lora((self.proj_out,), model_config.get("latent_lora", False)):
+        hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
+        gate = gate.unsqueeze(1)
+        hidden_states = gate * self.proj_out(hidden_states)
+        hidden_states = residual + hidden_states
+    if using_cond:
+        condition_latents = torch.cat([cond_attn_output, mlp_cond_hidden_states], dim=2)
+        cond_gate = cond_gate.unsqueeze(1)
+        condition_latents = cond_gate * self.proj_out(condition_latents)
+        condition_latents = residual_cond + condition_latents
+
+        extra_condition_latents = torch.cat([extra_cond_attn_output, extra_mlp_cond_hidden_states], dim=2)
+        extra_cond_gate = extra_cond_gate.unsqueeze(1)
+        extra_condition_latents = extra_cond_gate * self.proj_out(extra_condition_latents)
+        extra_condition_latents = extra_residual_cond + extra_condition_latents
+
+    if hidden_states.dtype == torch.float16:
+        hidden_states = hidden_states.clip(-65504, 65504)
+
+    return hidden_states if not using_cond else (hidden_states, condition_latents, extra_condition_latents)

flux/condition.py

@@ -0,0 +1,89 @@
+# Recycled from Ominicontrol and modified to accept an extra condition.
+# While Zenctrl pursued a similar idea, it diverged structurally. 
+# We appreciate the clarity of Omini's implementation and decided to align with it.
+
+import torch
+from typing import Optional, Union, List, Tuple
+from diffusers.pipelines import FluxPipeline
+from PIL import Image, ImageFilter
+import numpy as np
+import cv2
+
+# from pipeline_tools import encode_images
+from .pipeline_tools import encode_images
+
+condition_dict = {
+    "subject": 1,
+    "sr": 2,
+    "cot": 3,
+}
+
+
+class Condition(object):
+    def __init__(
+        self,
+        condition_type: str,
+        raw_img: Union[Image.Image, torch.Tensor] = None,
+        condition: Union[Image.Image, torch.Tensor] = None,
+        position_delta=None,
+    ) -> None:
+        self.condition_type = condition_type
+        assert raw_img is not None or condition is not None
+        if raw_img is not None:
+            self.condition = self.get_condition(condition_type, raw_img)
+        else:
+            self.condition = condition
+        self.position_delta = position_delta
+
+        
+    def get_condition(
+        self, condition_type: str, raw_img: Union[Image.Image, torch.Tensor]
+    ) -> Union[Image.Image, torch.Tensor]:
+        """
+        Returns the condition image.
+        """
+        if condition_type == "subject":
+            return raw_img
+        elif condition_type == "sr":
+            return raw_img
+        elif condition_type == "cot":
+            return raw_img.convert("RGB")
+        return self.condition
+
+    
+    @property
+    def type_id(self) -> int:
+        """
+        Returns the type id of the condition.
+        """
+        return condition_dict[self.condition_type]
+    
+    def encode(
+        self, pipe: FluxPipeline, empty: bool = False
+    ) -> Tuple[torch.Tensor, torch.Tensor, int]:
+        """
+        Encodes the condition into tokens, ids and type_id.
+        """
+        if self.condition_type in [
+            "subject",
+            "sr",
+            "cot"
+        ]:
+            if empty:
+                # make the condition black
+                e_condition = Image.new("RGB", self.condition.size, (0, 0, 0))
+                e_condition = e_condition.convert("RGB")
+                tokens, ids = encode_images(pipe, e_condition)
+            else:
+                tokens, ids = encode_images(pipe, self.condition)
+        else:
+            raise NotImplementedError(
+                f"Condition type {self.condition_type} not implemented"
+            )
+        if self.position_delta is None and self.condition_type == "subject":
+            self.position_delta = [0, -self.condition.size[0] // 16]
+        if self.position_delta is not None:
+            ids[:, 1] += self.position_delta[0]
+            ids[:, 2] += self.position_delta[1]
+        type_id = torch.ones_like(ids[:, :1]) * self.type_id
+        return tokens, ids, type_id

flux/generate.py

@@ -0,0 +1,337 @@
+# Recycled from Ominicontrol and modified to accept an extra condition.
+# While Zenctrl pursued a similar idea, it diverged structurally. 
+# We appreciate the clarity of Omini's implementation and decided to align with it.
+
+import torch
+import yaml, os
+from diffusers.pipelines import FluxPipeline
+from typing import List, Union, Optional, Dict, Any, Callable
+from .transformer import tranformer_forward
+from .condition import Condition
+
+
+from diffusers.pipelines.flux.pipeline_flux import (
+    FluxPipelineOutput,
+    calculate_shift,
+    retrieve_timesteps,
+    np,
+)
+
+
+def get_config(config_path: str = None):
+    config_path = config_path or os.environ.get("XFL_CONFIG")
+    if not config_path:
+        return {}
+    with open(config_path, "r") as f:
+        config = yaml.safe_load(f)
+    return config
+
+
+def prepare_params(
+    prompt: Union[str, List[str]] = None,
+    prompt_2: Optional[Union[str, List[str]]] = None,
+    height: Optional[int] = 512,
+    width: Optional[int] = 512,
+    num_inference_steps: int = 28,
+    timesteps: List[int] = None,
+    guidance_scale: float = 3.5,
+    num_images_per_prompt: Optional[int] = 1,
+    generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+    latents: Optional[torch.FloatTensor] = None,
+    prompt_embeds: Optional[torch.FloatTensor] = None,
+    pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+    output_type: Optional[str] = "pil",
+    return_dict: bool = True,
+    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+    callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+    max_sequence_length: int = 512,
+    **kwargs: dict,
+):
+    return (
+        prompt,
+        prompt_2,
+        height,
+        width,
+        num_inference_steps,
+        timesteps,
+        guidance_scale,
+        num_images_per_prompt,
+        generator,
+        latents,
+        prompt_embeds,
+        pooled_prompt_embeds,
+        output_type,
+        return_dict,
+        joint_attention_kwargs,
+        callback_on_step_end,
+        callback_on_step_end_tensor_inputs,
+        max_sequence_length,
+    )
+
+
+def seed_everything(seed: int = 42):
+    torch.backends.cudnn.deterministic = True
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+
+
+@torch.no_grad()
+def generate(
+    pipeline: FluxPipeline,
+    conditions: List[Condition] = None,
+    config_path: str = None,
+    model_config: Optional[Dict[str, Any]] = {},
+    condition_scale: float = [1, 1],
+    default_lora: bool = False,
+    image_guidance_scale: float = 1.0,
+    **params: dict,
+):
+    model_config = model_config or get_config(config_path).get("model", {})
+    if condition_scale != [1,1]:
+        for name, module in pipeline.transformer.named_modules():
+            if not name.endswith(".attn"):
+                continue
+            module.c_factor = torch.tensor(condition_scale) 
+
+    self = pipeline
+    (
+        prompt,
+        prompt_2,
+        height,
+        width,
+        num_inference_steps,
+        timesteps,
+        guidance_scale,
+        num_images_per_prompt,
+        generator,
+        latents,
+        prompt_embeds,
+        pooled_prompt_embeds,
+        output_type,
+        return_dict,
+        joint_attention_kwargs,
+        callback_on_step_end,
+        callback_on_step_end_tensor_inputs,
+        max_sequence_length,
+    ) = prepare_params(**params)
+
+    height = height or self.default_sample_size * self.vae_scale_factor
+    width = width or self.default_sample_size * self.vae_scale_factor
+
+    # 1. Check inputs. Raise error if not correct
+    self.check_inputs(
+        prompt,
+        prompt_2,
+        height,
+        width,
+        prompt_embeds=prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
+        callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+        max_sequence_length=max_sequence_length,
+    )
+
+    self._guidance_scale = guidance_scale
+    self._joint_attention_kwargs = joint_attention_kwargs
+    self._interrupt = False
+
+    # 2. Define call parameters
+    if prompt is not None and isinstance(prompt, str):
+        batch_size = 1
+    elif prompt is not None and isinstance(prompt, list):
+        batch_size = len(prompt)
+    else:
+        batch_size = prompt_embeds.shape[0]
+
+    device = self._execution_device
+
+    lora_scale = (
+        self.joint_attention_kwargs.get("scale", None)
+        if self.joint_attention_kwargs is not None
+        else None
+    )
+    (
+        prompt_embeds,
+        pooled_prompt_embeds,
+        text_ids,
+    ) = self.encode_prompt(
+        prompt=prompt,
+        prompt_2=prompt_2,
+        prompt_embeds=prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
+        device=device,
+        num_images_per_prompt=num_images_per_prompt,
+        max_sequence_length=max_sequence_length,
+        lora_scale=lora_scale,
+    )
+
+    # 4. Prepare latent variables
+    num_channels_latents = self.transformer.config.in_channels // 4
+    latents, latent_image_ids = self.prepare_latents(
+        batch_size * num_images_per_prompt,
+        num_channels_latents,
+        height,
+        width,
+        prompt_embeds.dtype,
+        device,
+        generator,
+        latents,
+    )
+
+    # 4.1. Prepare conditions
+    condition_latents, condition_ids, condition_type_ids = ([] for _ in range(3))
+    extra_condition_latents, extra_condition_ids, extra_condition_type_ids = ([] for _ in range(3))
+    use_condition = conditions is not None or []
+    if use_condition:
+        if not default_lora:
+            pipeline.set_adapters(conditions[1].condition_type)
+        # for condition in conditions:
+        tokens, ids, type_id = conditions[0].encode(self)
+        condition_latents.append(tokens)  # [batch_size, token_n, token_dim]
+        condition_ids.append(ids)  # [token_n, id_dim(3)]
+        condition_type_ids.append(type_id)  # [token_n, 1]
+        condition_latents = torch.cat(condition_latents, dim=1)
+        condition_ids = torch.cat(condition_ids, dim=0)
+        condition_type_ids = torch.cat(condition_type_ids, dim=0)
+
+        tokens, ids, type_id = conditions[1].encode(self)
+        extra_condition_latents.append(tokens)  # [batch_size, token_n, token_dim]
+        extra_condition_ids.append(ids)  # [token_n, id_dim(3)]
+        extra_condition_type_ids.append(type_id)  # [token_n, 1]
+        extra_condition_latents = torch.cat(extra_condition_latents, dim=1)
+        extra_condition_ids = torch.cat(extra_condition_ids, dim=0)
+        extra_condition_type_ids = torch.cat(extra_condition_type_ids, dim=0)
+
+    # 5. Prepare timesteps
+    sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+    image_seq_len = latents.shape[1]
+    mu = calculate_shift(
+        image_seq_len,
+        self.scheduler.config.base_image_seq_len,
+        self.scheduler.config.max_image_seq_len,
+        self.scheduler.config.base_shift,
+        self.scheduler.config.max_shift,
+    )
+    timesteps, num_inference_steps = retrieve_timesteps(
+        self.scheduler,
+        num_inference_steps,
+        device,
+        timesteps,
+        sigmas,
+        mu=mu,
+    )
+    num_warmup_steps = max(
+        len(timesteps) - num_inference_steps * self.scheduler.order, 0
+    )
+    self._num_timesteps = len(timesteps)
+
+    # 6. Denoising loop
+    with self.progress_bar(total=num_inference_steps) as progress_bar:
+        for i, t in enumerate(timesteps):
+            if self.interrupt:
+                continue
+
+            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+            timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+            # handle guidance
+            if self.transformer.config.guidance_embeds:
+                guidance = torch.tensor([guidance_scale], device=device)
+                guidance = guidance.expand(latents.shape[0])
+            else:
+                guidance = None
+            noise_pred = tranformer_forward(
+                self.transformer,
+                model_config=model_config,
+                # Inputs of the condition (new feature)
+                condition_latents=condition_latents if use_condition else None,
+                condition_ids=condition_ids if use_condition else None,
+                condition_type_ids=condition_type_ids if use_condition else None,
+                extra_condition_latents=extra_condition_latents if use_condition else None,
+                extra_condition_ids=extra_condition_ids if use_condition else None,
+                extra_condition_type_ids=extra_condition_type_ids if use_condition else None,
+                # Inputs to the original transformer
+                hidden_states=latents,
+                # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                timestep=timestep / 1000,
+                guidance=guidance,
+                pooled_projections=pooled_prompt_embeds,
+                encoder_hidden_states=prompt_embeds,
+                txt_ids=text_ids,
+                img_ids=latent_image_ids,
+                joint_attention_kwargs=self.joint_attention_kwargs,
+                return_dict=False,
+            )[0]
+
+            if image_guidance_scale != 1.0:
+                uncondition_latents = conditions.encode(self, empty=True)[0]
+                unc_pred = tranformer_forward(
+                    self.transformer,
+                    model_config=model_config,
+                    # Inputs of the condition (new feature)
+                    condition_latents=uncondition_latents if use_condition else None,
+                    condition_ids=condition_ids if use_condition else None,
+                    condition_type_ids=condition_type_ids if use_condition else None,
+                    # Inputs to the original transformer
+                    hidden_states=latents,
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                    timestep=timestep / 1000,
+                    guidance=torch.ones_like(guidance),
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                noise_pred = unc_pred + image_guidance_scale * (noise_pred - unc_pred)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents_dtype = latents.dtype
+            latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+            if latents.dtype != latents_dtype:
+                if torch.backends.mps.is_available():
+                    # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                    latents = latents.to(latents_dtype)
+
+            if callback_on_step_end is not None:
+                callback_kwargs = {}
+                for k in callback_on_step_end_tensor_inputs:
+                    callback_kwargs[k] = locals()[k]
+                callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                latents = callback_outputs.pop("latents", latents)
+                prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+            # call the callback, if provided
+            if i == len(timesteps) - 1 or (
+                (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+            ):
+                progress_bar.update()
+
+    if output_type == "latent":
+        image = latents
+
+    else:
+        latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+        latents = (
+            latents / self.vae.config.scaling_factor
+        ) + self.vae.config.shift_factor
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = self.image_processor.postprocess(image, output_type=output_type)
+
+    # Offload all models
+    self.maybe_free_model_hooks()
+
+    if condition_scale != [1,1]:
+        for name, module in pipeline.transformer.named_modules():
+            if not name.endswith(".attn"):
+                continue
+            del module.c_factor
+
+    if not return_dict:
+        return (image,)
+
+    return FluxPipelineOutput(images=image)

flux/lora_controller.py

@@ -0,0 +1,82 @@
+#As is from OminiControl
+from peft.tuners.tuners_utils import BaseTunerLayer
+from typing import List, Any, Optional, Type
+from .condition import condition_dict
+
+
+class enable_lora:
+    def __init__(self, lora_modules: List[BaseTunerLayer], activated: bool) -> None:
+        self.activated: bool = activated
+        if activated:
+            return
+        self.lora_modules: List[BaseTunerLayer] = [
+            each for each in lora_modules if isinstance(each, BaseTunerLayer)
+        ]
+        self.scales = [
+            {
+                active_adapter: lora_module.scaling[active_adapter]
+                for active_adapter in lora_module.active_adapters
+            }
+            for lora_module in self.lora_modules
+        ]
+
+    def __enter__(self) -> None:
+        if self.activated:
+            return
+
+        for lora_module in self.lora_modules:
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            for active_adapter in lora_module.active_adapters:
+                if (
+                    active_adapter in condition_dict.keys()
+                    or active_adapter == "default"
+                ):
+                    lora_module.scaling[active_adapter] = 0.0
+
+    def __exit__(
+        self,
+        exc_type: Optional[Type[BaseException]],
+        exc_val: Optional[BaseException],
+        exc_tb: Optional[Any],
+    ) -> None:
+        if self.activated:
+            return
+        for i, lora_module in enumerate(self.lora_modules):
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            for active_adapter in lora_module.active_adapters:
+                lora_module.scaling[active_adapter] = self.scales[i][active_adapter]
+
+
+class set_lora_scale:
+    def __init__(self, lora_modules: List[BaseTunerLayer], scale: float) -> None:
+        self.lora_modules: List[BaseTunerLayer] = [
+            each for each in lora_modules if isinstance(each, BaseTunerLayer)
+        ]
+        self.scales = [
+            {
+                active_adapter: lora_module.scaling[active_adapter]
+                for active_adapter in lora_module.active_adapters
+            }
+            for lora_module in self.lora_modules
+        ]
+        self.scale = scale
+
+    def __enter__(self) -> None:
+        for lora_module in self.lora_modules:
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            lora_module.scale_layer(self.scale)
+
+    def __exit__(
+        self,
+        exc_type: Optional[Type[BaseException]],
+        exc_val: Optional[BaseException],
+        exc_tb: Optional[Any],
+    ) -> None:
+        for i, lora_module in enumerate(self.lora_modules):
+            if not isinstance(lora_module, BaseTunerLayer):
+                continue
+            for active_adapter in lora_module.active_adapters:
+                lora_module.scaling[active_adapter] = self.scales[i][active_adapter]

flux/pipeline_tools.py

@@ -0,0 +1,53 @@
+#As is from OminiControl
+from diffusers.pipelines import FluxPipeline
+from diffusers.utils import logging
+from diffusers.pipelines.flux.pipeline_flux import logger
+from torch import Tensor
+
+
+def encode_images(pipeline: FluxPipeline, images: Tensor):
+    images = pipeline.image_processor.preprocess(images)
+    images = images.to(pipeline.device).to(pipeline.dtype)
+    images = pipeline.vae.encode(images).latent_dist.sample()
+    images = (
+        images - pipeline.vae.config.shift_factor
+    ) * pipeline.vae.config.scaling_factor
+    images_tokens = pipeline._pack_latents(images, *images.shape)
+    images_ids = pipeline._prepare_latent_image_ids(
+        images.shape[0],
+        images.shape[2],
+        images.shape[3],
+        pipeline.device,
+        pipeline.dtype,
+    )
+    if images_tokens.shape[1] != images_ids.shape[0]:
+        images_ids = pipeline._prepare_latent_image_ids(
+            images.shape[0],
+            images.shape[2] // 2,
+            images.shape[3] // 2,
+            pipeline.device,
+            pipeline.dtype,
+        )
+    return images_tokens, images_ids
+
+
+def prepare_text_input(pipeline: FluxPipeline, prompts, max_sequence_length=512):
+    # Turn off warnings (CLIP overflow)
+    logger.setLevel(logging.ERROR)
+    (
+        prompt_embeds,
+        pooled_prompt_embeds,
+        text_ids,
+    ) = pipeline.encode_prompt(
+        prompt=prompts,
+        prompt_2=None,
+        prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        device=pipeline.device,
+        num_images_per_prompt=1,
+        max_sequence_length=max_sequence_length,
+        lora_scale=None,
+    )
+    # Turn on warnings
+    logger.setLevel(logging.WARNING)
+    return prompt_embeds, pooled_prompt_embeds, text_ids

flux/transformer.py

@@ -0,0 +1,286 @@
+# Recycled from Ominicontrol and modified to accept an extra condition.
+# While Zenctrl pursued a similar idea, it diverged structurally. 
+# We appreciate the clarity of Omini's implementation and decided to align with it.
+
+import torch
+from diffusers.pipelines import FluxPipeline
+from typing import List, Union, Optional, Dict, Any, Callable
+from .block import block_forward, single_block_forward
+from .lora_controller import enable_lora
+from accelerate.utils import is_torch_version
+from diffusers.models.transformers.transformer_flux import (
+    FluxTransformer2DModel,
+    Transformer2DModelOutput,
+    USE_PEFT_BACKEND,
+    scale_lora_layers,
+    unscale_lora_layers,
+    logger,
+)
+import numpy as np
+
+
+def prepare_params(
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor = None,
+    pooled_projections: torch.Tensor = None,
+    timestep: torch.LongTensor = None,
+    img_ids: torch.Tensor = None,
+    txt_ids: torch.Tensor = None,
+    guidance: torch.Tensor = None,
+    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    controlnet_block_samples=None,
+    controlnet_single_block_samples=None,
+    return_dict: bool = True,
+    **kwargs: dict,
+):
+    return (
+        hidden_states,
+        encoder_hidden_states,
+        pooled_projections,
+        timestep,
+        img_ids,
+        txt_ids,
+        guidance,
+        joint_attention_kwargs,
+        controlnet_block_samples,
+        controlnet_single_block_samples,
+        return_dict,
+    )
+
+
+def tranformer_forward(
+    transformer: FluxTransformer2DModel,
+    condition_latents: torch.Tensor,
+    extra_condition_latents: torch.Tensor,
+    condition_ids: torch.Tensor,
+    condition_type_ids: torch.Tensor,
+    extra_condition_ids: torch.Tensor,
+    extra_condition_type_ids: torch.Tensor,
+    model_config: Optional[Dict[str, Any]] = {},
+    c_t=0,
+    **params: dict,
+):
+    self = transformer
+    use_condition = condition_latents is not None
+    use_extra_condition = extra_condition_latents is not None
+
+    (
+        hidden_states,
+        encoder_hidden_states,
+        pooled_projections,
+        timestep,
+        img_ids,
+        txt_ids,
+        guidance,
+        joint_attention_kwargs,
+        controlnet_block_samples,
+        controlnet_single_block_samples,
+        return_dict,
+    ) = prepare_params(**params)
+
+    if joint_attention_kwargs is not None:
+        joint_attention_kwargs = joint_attention_kwargs.copy()
+        lora_scale = joint_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)
+    else:
+        if (
+            joint_attention_kwargs is not None
+            and joint_attention_kwargs.get("scale", None) is not None
+        ):
+            logger.warning(
+                "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+            )
+
+    with enable_lora((self.x_embedder,), model_config.get("latent_lora", False)):
+        hidden_states = self.x_embedder(hidden_states)
+    condition_latents = self.x_embedder(condition_latents) if use_condition else None
+    extra_condition_latents = self.x_embedder(extra_condition_latents) if use_extra_condition else None
+
+    timestep = timestep.to(hidden_states.dtype) * 1000
+
+    if guidance is not None:
+        guidance = guidance.to(hidden_states.dtype) * 1000
+    else:
+        guidance = None
+
+    temb = (
+        self.time_text_embed(timestep, pooled_projections)
+        if guidance is None
+        else self.time_text_embed(timestep, guidance, pooled_projections)
+    )
+
+    cond_temb = (
+        self.time_text_embed(torch.ones_like(timestep) * c_t * 1000, pooled_projections)
+        if guidance is None
+        else self.time_text_embed(
+            torch.ones_like(timestep) * c_t * 1000, torch.ones_like(guidance) * 1000, pooled_projections
+        )
+    )
+    extra_cond_temb = (
+        self.time_text_embed(torch.ones_like(timestep) * c_t * 1000, pooled_projections)
+        if guidance is None
+        else self.time_text_embed(
+            torch.ones_like(timestep) * c_t * 1000, torch.ones_like(guidance) * 1000, pooled_projections
+        )
+    )
+    encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+    if txt_ids.ndim == 3:
+        logger.warning(
+            "Passing `txt_ids` 3d torch.Tensor is deprecated."
+            "Please remove the batch dimension and pass it as a 2d torch Tensor"
+        )
+        txt_ids = txt_ids[0]
+    if img_ids.ndim == 3:
+        logger.warning(
+            "Passing `img_ids` 3d torch.Tensor is deprecated."
+            "Please remove the batch dimension and pass it as a 2d torch Tensor"
+        )
+        img_ids = img_ids[0]
+
+    ids = torch.cat((txt_ids, img_ids), dim=0)
+    image_rotary_emb = self.pos_embed(ids)
+    if use_condition:
+        # condition_ids[:, :1] = condition_type_ids
+        cond_rotary_emb = self.pos_embed(condition_ids)
+
+    if use_extra_condition:
+        extra_cond_rotary_emb = self.pos_embed(extra_condition_ids)
+
+
+    # hidden_states = torch.cat([hidden_states, condition_latents], dim=1)
+
+    #print("here!")
+    for index_block, block in enumerate(self.transformer_blocks):
+        if self.training and self.gradient_checkpointing:
+            ckpt_kwargs: Dict[str, Any] = (
+                {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+            )
+            encoder_hidden_states, hidden_states, condition_latents, extra_condition_latents = (
+                torch.utils.checkpoint.checkpoint(
+                    block_forward,
+                    self=block,
+                    model_config=model_config,
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    condition_latents=condition_latents if use_condition else None,
+                    extra_condition_latents=extra_condition_latents if use_extra_condition else None,
+                    temb=temb,
+                    cond_temb=cond_temb if use_condition else None,
+                    cond_rotary_emb=cond_rotary_emb if use_condition else None,
+                    extra_cond_temb=extra_cond_temb if use_extra_condition else None,
+                    extra_cond_rotary_emb=extra_cond_rotary_emb if use_extra_condition else None,
+                    image_rotary_emb=image_rotary_emb,
+                    **ckpt_kwargs,
+                )
+            )
+
+        else:
+            encoder_hidden_states, hidden_states, condition_latents, extra_condition_latents = block_forward(
+                block,
+                model_config=model_config,
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                condition_latents=condition_latents if use_condition else None,
+                extra_condition_latents=extra_condition_latents if use_extra_condition else None,
+                temb=temb,
+                cond_temb=cond_temb if use_condition else None,
+                cond_rotary_emb=cond_rotary_emb if use_condition else None,
+                extra_cond_temb=cond_temb if use_extra_condition else None,
+                extra_cond_rotary_emb=extra_cond_rotary_emb if use_extra_condition else None,
+                image_rotary_emb=image_rotary_emb,
+            )
+
+        # controlnet residual
+        if controlnet_block_samples is not None:
+            interval_control = len(self.transformer_blocks) / len(
+                controlnet_block_samples
+            )
+            interval_control = int(np.ceil(interval_control))
+            hidden_states = (
+                hidden_states
+                + controlnet_block_samples[index_block // interval_control]
+            )
+    hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+    
+
+    for index_block, block in enumerate(self.single_transformer_blocks):
+        if self.training and self.gradient_checkpointing:
+            ckpt_kwargs: Dict[str, Any] = (
+                {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+            )
+            result = torch.utils.checkpoint.checkpoint(
+                single_block_forward,
+                self=block,
+                model_config=model_config,
+                hidden_states=hidden_states,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+                **(
+                    {
+                        "condition_latents": condition_latents,
+                        "extra_condition_latents": extra_condition_latents,
+                        "cond_temb": cond_temb,
+                        "cond_rotary_emb": cond_rotary_emb,
+                        "extra_cond_temb": extra_cond_temb,
+                        "extra_cond_rotary_emb": extra_cond_rotary_emb,
+                    }
+                    if use_condition
+                    else {}
+                ),
+                **ckpt_kwargs,
+            )
+
+        else:
+            result = single_block_forward(
+                block,
+                model_config=model_config,
+                hidden_states=hidden_states,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+                **(
+                    {
+                        "condition_latents": condition_latents,
+                        "extra_condition_latents": extra_condition_latents,
+                        "cond_temb": cond_temb,
+                        "cond_rotary_emb": cond_rotary_emb,
+                        "extra_cond_temb": extra_cond_temb,
+                        "extra_cond_rotary_emb": extra_cond_rotary_emb,
+                    }
+                    if use_condition
+                    else {}
+                ),
+            )
+        if use_condition:
+            hidden_states, condition_latents, extra_condition_latents = result
+        else:
+            hidden_states = result
+
+        # controlnet residual
+        if controlnet_single_block_samples is not None:
+            interval_control = len(self.single_transformer_blocks) / len(
+                controlnet_single_block_samples
+            )
+            interval_control = int(np.ceil(interval_control))
+            hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
+                hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+                + controlnet_single_block_samples[index_block // interval_control]
+            )
+
+    hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+
+    hidden_states = self.norm_out(hidden_states, temb)
+    output = self.proj_out(hidden_states)
+
+    if USE_PEFT_BACKEND:
+        # remove `lora_scale` from each PEFT layer
+        unscale_lora_layers(self, lora_scale)
+
+    if not return_dict:
+        return (output,)
+    return Transformer2DModelOutput(sample=output)