diffusers 0.29.0__py3-none-any.whl → 0.29.1__py3-none-any.whl

diffusers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.29.0"
+__version__ = "0.29.1"
 
 from typing import TYPE_CHECKING
 
@@ -91,6 +91,8 @@ else:
             "MultiAdapter",
             "PixArtTransformer2DModel",
             "PriorTransformer",
+            "SD3ControlNetModel",
+            "SD3MultiControlNetModel",
             "SD3Transformer2DModel",
             "StableCascadeUNet",
             "T2IAdapter",
@@ -278,6 +280,7 @@ else:
             "StableCascadeCombinedPipeline",
             "StableCascadeDecoderPipeline",
             "StableCascadePriorPipeline",
+            "StableDiffusion3ControlNetPipeline",
             "StableDiffusion3Img2ImgPipeline",
             "StableDiffusion3Pipeline",
             "StableDiffusionAdapterPipeline",
@@ -501,6 +504,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             MultiAdapter,
             PixArtTransformer2DModel,
             PriorTransformer,
+            SD3ControlNetModel,
+            SD3MultiControlNetModel,
             SD3Transformer2DModel,
             T2IAdapter,
             T5FilmDecoder,
@@ -666,6 +671,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableCascadeCombinedPipeline,
             StableCascadeDecoderPipeline,
             StableCascadePriorPipeline,
+            StableDiffusion3ControlNetPipeline,
             StableDiffusion3Img2ImgPipeline,
             StableDiffusion3Pipeline,
             StableDiffusionAdapterPipeline,

diffusers/loaders/single_file.py

@@ -28,9 +28,11 @@ from .single_file_utils import (
     _legacy_load_safety_checker,
     _legacy_load_scheduler,
     create_diffusers_clip_model_from_ldm,
+    create_diffusers_t5_model_from_checkpoint,
     fetch_diffusers_config,
     fetch_original_config,
     is_clip_model_in_single_file,
+    is_t5_in_single_file,
     load_single_file_checkpoint,
 )
 
@@ -118,6 +120,16 @@ def load_single_file_sub_model(
             is_legacy_loading=is_legacy_loading,
         )
 
+    elif is_transformers_model and is_t5_in_single_file(checkpoint):
+        loaded_sub_model = create_diffusers_t5_model_from_checkpoint(
+            class_obj,
+            checkpoint=checkpoint,
+            config=cached_model_config_path,
+            subfolder=name,
+            torch_dtype=torch_dtype,
+            local_files_only=local_files_only,
+        )
+
     elif is_tokenizer and is_legacy_loading:
         loaded_sub_model = _legacy_load_clip_tokenizer(
             class_obj, checkpoint=checkpoint, config=cached_model_config_path, local_files_only=local_files_only

diffusers/loaders/single_file_model.py

@@ -276,16 +276,18 @@ class FromOriginalModelMixin:
 
         if is_accelerate_available():
             unexpected_keys = load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
-            if model._keys_to_ignore_on_load_unexpected is not None:
-                for pat in model._keys_to_ignore_on_load_unexpected:
-                    unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
 
-            if len(unexpected_keys) > 0:
-                logger.warning(
-                    f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
-                )
         else:
-            model.load_state_dict(diffusers_format_checkpoint)
+            _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
+
+        if model._keys_to_ignore_on_load_unexpected is not None:
+            for pat in model._keys_to_ignore_on_load_unexpected:
+                unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+        if len(unexpected_keys) > 0:
+            logger.warning(
+                f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
+            )
 
         if torch_dtype is not None:
             model.to(torch_dtype)

diffusers/loaders/single_file_utils.py

@@ -252,7 +252,6 @@ LDM_CONTROLNET_KEY = "control_model."
 LDM_CLIP_PREFIX_TO_REMOVE = [
     "cond_stage_model.transformer.",
     "conditioner.embedders.0.transformer.",
-    "text_encoders.clip_l.transformer.",
 ]
 OPEN_CLIP_PREFIX = "conditioner.embedders.0.model."
 LDM_OPEN_CLIP_TEXT_PROJECTION_DIM = 1024
@@ -399,11 +398,14 @@ def is_open_clip_sdxl_model(checkpoint):
 
 
 def is_open_clip_sd3_model(checkpoint):
-    is_open_clip_sdxl_refiner_model(checkpoint)
+    if CHECKPOINT_KEY_NAMES["open_clip_sd3"] in checkpoint:
+        return True
+
+    return False
 
 
 def is_open_clip_sdxl_refiner_model(checkpoint):
-    if CHECKPOINT_KEY_NAMES["open_clip_sd3"] in checkpoint:
+    if CHECKPOINT_KEY_NAMES["open_clip_sdxl_refiner"] in checkpoint:
         return True
 
     return False
@@ -1233,11 +1235,14 @@ def convert_ldm_vae_checkpoint(checkpoint, config):
     return new_checkpoint
 
 
-def convert_ldm_clip_checkpoint(checkpoint):
+def convert_ldm_clip_checkpoint(checkpoint, remove_prefix=None):
     keys = list(checkpoint.keys())
     text_model_dict = {}
 
-    remove_prefixes = LDM_CLIP_PREFIX_TO_REMOVE
+    remove_prefixes = []
+    remove_prefixes.extend(LDM_CLIP_PREFIX_TO_REMOVE)
+    if remove_prefix:
+        remove_prefixes.append(remove_prefix)
 
     for key in keys:
         for prefix in remove_prefixes:
@@ -1263,8 +1268,6 @@ def convert_open_clip_checkpoint(
     else:
         text_proj_dim = LDM_OPEN_CLIP_TEXT_PROJECTION_DIM
 
-    text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids")
-
     keys = list(checkpoint.keys())
     keys_to_ignore = SD_2_TEXT_ENCODER_KEYS_TO_IGNORE
 
@@ -1313,9 +1316,6 @@ def convert_open_clip_checkpoint(
         else:
             text_model_dict[diffusers_key] = checkpoint.get(key)
 
-    if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)):
-        text_model_dict.pop("text_model.embeddings.position_ids", None)
-
     return text_model_dict
 
 
@@ -1376,6 +1376,13 @@ def create_diffusers_clip_model_from_ldm(
     ):
         diffusers_format_checkpoint = convert_ldm_clip_checkpoint(checkpoint)
 
+    elif (
+        is_clip_sd3_model(checkpoint)
+        and checkpoint[CHECKPOINT_KEY_NAMES["clip_sd3"]].shape[-1] == position_embedding_dim
+    ):
+        diffusers_format_checkpoint = convert_ldm_clip_checkpoint(checkpoint, "text_encoders.clip_l.transformer.")
+        diffusers_format_checkpoint["text_projection.weight"] = torch.eye(position_embedding_dim)
+
     elif is_open_clip_model(checkpoint):
         prefix = "cond_stage_model.model."
         diffusers_format_checkpoint = convert_open_clip_checkpoint(model, checkpoint, prefix=prefix)
@@ -1391,26 +1398,28 @@ def create_diffusers_clip_model_from_ldm(
         prefix = "conditioner.embedders.0.model."
         diffusers_format_checkpoint = convert_open_clip_checkpoint(model, checkpoint, prefix=prefix)
 
-    elif is_open_clip_sd3_model(checkpoint):
-        prefix = "text_encoders.clip_g.transformer."
-        diffusers_format_checkpoint = convert_open_clip_checkpoint(model, checkpoint, prefix=prefix)
+    elif (
+        is_open_clip_sd3_model(checkpoint)
+        and checkpoint[CHECKPOINT_KEY_NAMES["open_clip_sd3"]].shape[-1] == position_embedding_dim
+    ):
+        diffusers_format_checkpoint = convert_ldm_clip_checkpoint(checkpoint, "text_encoders.clip_g.transformer.")
 
     else:
         raise ValueError("The provided checkpoint does not seem to contain a valid CLIP model.")
 
     if is_accelerate_available():
         unexpected_keys = load_model_dict_into_meta(model, diffusers_format_checkpoint, dtype=torch_dtype)
-        if model._keys_to_ignore_on_load_unexpected is not None:
-            for pat in model._keys_to_ignore_on_load_unexpected:
-                unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+    else:
+        _, unexpected_keys = model.load_state_dict(diffusers_format_checkpoint, strict=False)
 
-        if len(unexpected_keys) > 0:
-            logger.warning(
-                f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
-            )
+    if model._keys_to_ignore_on_load_unexpected is not None:
+        for pat in model._keys_to_ignore_on_load_unexpected:
+            unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
 
-    else:
-        model.load_state_dict(diffusers_format_checkpoint)
+    if len(unexpected_keys) > 0:
+        logger.warning(
+            f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
+        )
 
     if torch_dtype is not None:
         model.to(torch_dtype)
@@ -1755,7 +1764,7 @@ def convert_sd3_t5_checkpoint_to_diffusers(checkpoint):
     keys = list(checkpoint.keys())
     text_model_dict = {}
 
-    remove_prefixes = ["text_encoders.t5xxl.transformer.encoder."]
+    remove_prefixes = ["text_encoders.t5xxl.transformer."]
 
     for key in keys:
         for prefix in remove_prefixes:
@@ -1799,3 +1808,4 @@ def create_diffusers_t5_model_from_checkpoint(
 
     else:
         model.load_state_dict(diffusers_format_checkpoint)
+    return model

diffusers/models/__init__.py

@@ -33,6 +33,7 @@ if is_torch_available():
     _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["autoencoders.vq_model"] = ["VQModel"]
     _import_structure["controlnet"] = ["ControlNetModel"]
+    _import_structure["controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
@@ -74,6 +75,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VQModel,
         )
         from .controlnet import ControlNetModel
+        from .controlnet_sd3 import SD3ControlNetModel, SD3MultiControlNetModel
         from .controlnet_xs import ControlNetXSAdapter, UNetControlNetXSModel
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin

diffusers/models/controlnet_sd3.py

@@ -0,0 +1,418 @@
+# Copyright 2024 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ..models.attention import JointTransformerBlock
+from ..models.attention_processor import Attention, AttentionProcessor
+from ..models.modeling_utils import ModelMixin
+from ..utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from .controlnet import BaseOutput, zero_module
+from .embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
+from .transformers.transformer_2d import Transformer2DModelOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class SD3ControlNetOutput(BaseOutput):
+    controlnet_block_samples: Tuple[torch.Tensor]
+
+
+class SD3ControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: int = 128,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        num_layers: int = 18,
+        attention_head_dim: int = 64,
+        num_attention_heads: int = 18,
+        joint_attention_dim: int = 4096,
+        caption_projection_dim: int = 1152,
+        pooled_projection_dim: int = 2048,
+        out_channels: int = 16,
+        pos_embed_max_size: int = 96,
+    ):
+        super().__init__()
+        default_out_channels = in_channels
+        self.out_channels = out_channels if out_channels is not None else default_out_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pos_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=self.inner_dim,
+            pos_embed_max_size=pos_embed_max_size,
+        )
+        self.time_text_embed = CombinedTimestepTextProjEmbeddings(
+            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
+        )
+        self.context_embedder = nn.Linear(joint_attention_dim, caption_projection_dim)
+
+        # `attention_head_dim` is doubled to account for the mixing.
+        # It needs to crafted when we get the actual checkpoints.
+        self.transformer_blocks = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=self.inner_dim,
+                    context_pre_only=False,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        # controlnet_blocks
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            controlnet_block = nn.Linear(self.inner_dim, self.inner_dim)
+            controlnet_block = zero_module(controlnet_block)
+            self.controlnet_blocks.append(controlnet_block)
+        pos_embed_input = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=self.inner_dim,
+            pos_embed_type=None,
+        )
+        self.pos_embed_input = zero_module(pos_embed_input)
+
+        self.gradient_checkpointing = False
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+    def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+        """
+        Sets the attention processor to use [feed forward
+        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+        Parameters:
+            chunk_size (`int`, *optional*):
+                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+                over each tensor of dim=`dim`.
+            dim (`int`, *optional*, defaults to `0`):
+                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+                or dim=1 (sequence length).
+        """
+        if dim not in [0, 1]:
+            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+        # By default chunk size is 1
+        chunk_size = chunk_size or 1
+
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, chunk_size, dim)
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    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
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    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)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    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)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    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 _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    @classmethod
+    def from_transformer(cls, transformer, num_layers=None, load_weights_from_transformer=True):
+        config = transformer.config
+        config["num_layers"] = num_layers or config.num_layers
+        controlnet = cls(**config)
+
+        if load_weights_from_transformer:
+            controlnet.pos_embed.load_state_dict(transformer.pos_embed.state_dict(), strict=False)
+            controlnet.time_text_embed.load_state_dict(transformer.time_text_embed.state_dict(), strict=False)
+            controlnet.context_embedder.load_state_dict(transformer.context_embedder.state_dict(), strict=False)
+            controlnet.transformer_blocks.load_state_dict(transformer.transformer_blocks.state_dict())
+
+            controlnet.pos_embed_input = zero_module(controlnet.pos_embed_input)
+
+        return controlnet
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        controlnet_cond: torch.Tensor,
+        conditioning_scale: float = 1.0,
+        encoder_hidden_states: torch.FloatTensor = None,
+        pooled_projections: torch.FloatTensor = None,
+        timestep: torch.LongTensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
+        """
+        The [`SD3Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            controlnet_cond (`torch.Tensor`):
+                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
+                from the embeddings of input conditions.
+            timestep ( `torch.LongTensor`):
+                Used to indicate denoising step.
+            joint_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).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] 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.
+        """
+        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."
+                )
+
+        height, width = hidden_states.shape[-2:]
+
+        hidden_states = self.pos_embed(hidden_states)  # takes care of adding positional embeddings too.
+        temb = self.time_text_embed(timestep, pooled_projections)
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+        # add
+        hidden_states = hidden_states + self.pos_embed_input(controlnet_cond)
+
+        block_res_samples = ()
+
+        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,
+                    encoder_hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, temb=temb
+                )
+
+            block_res_samples = block_res_samples + (hidden_states,)
+
+        controlnet_block_res_samples = ()
+        for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
+            block_res_sample = controlnet_block(block_res_sample)
+            controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
+
+        # 6. scaling
+        controlnet_block_res_samples = [sample * conditioning_scale for sample in controlnet_block_res_samples]
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (controlnet_block_res_samples,)
+
+        return SD3ControlNetOutput(controlnet_block_samples=controlnet_block_res_samples)
+
+
+class SD3MultiControlNetModel(ModelMixin):
+    r"""
+    `SD3ControlNetModel` wrapper class for Multi-SD3ControlNet
+
+    This module is a wrapper for multiple instances of the `SD3ControlNetModel`. The `forward()` API is designed to be
+    compatible with `SD3ControlNetModel`.
+
+    Args:
+        controlnets (`List[SD3ControlNetModel]`):
+            Provides additional conditioning to the unet during the denoising process. You must set multiple
+            `SD3ControlNetModel` as a list.
+    """
+
+    def __init__(self, controlnets):
+        super().__init__()
+        self.nets = nn.ModuleList(controlnets)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        controlnet_cond: List[torch.tensor],
+        conditioning_scale: List[float],
+        pooled_projections: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        timestep: torch.LongTensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[SD3ControlNetOutput, Tuple]:
+        for i, (image, scale, controlnet) in enumerate(zip(controlnet_cond, conditioning_scale, self.nets)):
+            block_samples = controlnet(
+                hidden_states=hidden_states,
+                timestep=timestep,
+                encoder_hidden_states=encoder_hidden_states,
+                pooled_projections=pooled_projections,
+                controlnet_cond=image,
+                conditioning_scale=scale,
+                joint_attention_kwargs=joint_attention_kwargs,
+                return_dict=return_dict,
+            )
+
+            # merge samples
+            if i == 0:
+                control_block_samples = block_samples
+            else:
+                control_block_samples = [
+                    control_block_sample + block_sample
+                    for control_block_sample, block_sample in zip(control_block_samples[0], block_samples[0])
+                ]
+                control_block_samples = (tuple(control_block_samples),)
+
+        return control_block_samples