[Core] introduce `controlnet` module (#8768)

* move vae flax module.

* controlnet module.

* prepare for PR.

* revert a commit

* gracefully deprecate controlnet deps.

* fix

* fix doc path

* fix-copies

* fix path

* style

* style

* conflicts

* fix

* fix-copies

* sparsectrl.

* updates

* fix

* updates

* updates

* updates

* fix

---------
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

docs/source/en/api/models/controlnet.md

@@ -39,7 +39,7 @@ pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=contro
 
 ## ControlNetOutput
 
-[[autodoc]] models.controlnet.ControlNetOutput
+[[autodoc]] models.controlnets.controlnet.ControlNetOutput
 
 ## FlaxControlNetModel
 
@@ -47,4 +47,4 @@ pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=contro
 
 ## FlaxControlNetOutput
 
-[[autodoc]] models.controlnet_flax.FlaxControlNetOutput
+[[autodoc]] models.controlnets.controlnet_flax.FlaxControlNetOutput

docs/source/en/api/models/controlnet_sd3.md

@@ -38,5 +38,5 @@ pipe = StableDiffusion3ControlNetPipeline.from_pretrained("stabilityai/stable-di
 
 ## SD3ControlNetOutput
 
-[[autodoc]] models.controlnet_sd3.SD3ControlNetOutput
+[[autodoc]] models.controlnets.controlnet_sd3.SD3ControlNetOutput
 

examples/research_projects/promptdiffusion/promptdiffusioncontrolnet.py

@@ -229,11 +229,11 @@ class PromptDiffusionControlNetModel(ControlNetModel):
                 In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
                 you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
             return_dict (`bool`, defaults to `True`):
-                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+                Whether or not to return a [`~models.controlnets.controlnet.ControlNetOutput`] instead of a plain tuple.
 
         Returns:
-            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
-                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
+            [`~models.controlnets.controlnet.ControlNetOutput`] **or** `tuple`:
+                If `return_dict` is `True`, a [`~models.controlnets.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
                 returned where the first element is the sample tensor.
         """
         # check channel order

src/diffusers/__init__.py

@@ -487,7 +487,7 @@ except OptionalDependencyNotAvailable:
 
 
 else:
-    _import_structure["models.controlnet_flax"] = ["FlaxControlNetModel"]
+    _import_structure["models.controlnets.controlnet_flax"] = ["FlaxControlNetModel"]
     _import_structure["models.modeling_flax_utils"] = ["FlaxModelMixin"]
     _import_structure["models.unets.unet_2d_condition_flax"] = ["FlaxUNet2DConditionModel"]
     _import_structure["models.vae_flax"] = ["FlaxAutoencoderKL"]
@@ -914,7 +914,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from .utils.dummy_flax_objects import *  # noqa F403
     else:
-        from .models.controlnet_flax import FlaxControlNetModel
+        from .models.controlnets.controlnet_flax import FlaxControlNetModel
         from .models.modeling_flax_utils import FlaxModelMixin
         from .models.unets.unet_2d_condition_flax import FlaxUNet2DConditionModel
         from .models.vae_flax import FlaxAutoencoderKL

src/diffusers/models/__init__.py

@@ -36,12 +36,16 @@ if is_torch_available():
     _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
     _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
     _import_structure["autoencoders.vq_model"] = ["VQModel"]
-    _import_structure["controlnet"] = ["ControlNetModel"]
-    _import_structure["controlnet_flux"] = ["FluxControlNetModel", "FluxMultiControlNetModel"]
-    _import_structure["controlnet_hunyuan"] = ["HunyuanDiT2DControlNetModel", "HunyuanDiT2DMultiControlNetModel"]
-    _import_structure["controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
-    _import_structure["controlnet_sparsectrl"] = ["SparseControlNetModel"]
-    _import_structure["controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
+    _import_structure["controlnets.controlnet"] = ["ControlNetModel"]
+    _import_structure["controlnets.controlnet_flux"] = ["FluxControlNetModel", "FluxMultiControlNetModel"]
+    _import_structure["controlnets.controlnet_hunyuan"] = [
+        "HunyuanDiT2DControlNetModel",
+        "HunyuanDiT2DMultiControlNetModel",
+    ]
+    _import_structure["controlnets.controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
+    _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"]
+    _import_structure["controlnets.controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
+    _import_structure["controlnets.multicontrolnet"] = ["MultiControlNetModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
     _import_structure["modeling_utils"] = ["ModelMixin"]
     _import_structure["transformers.auraflow_transformer_2d"] = ["AuraFlowTransformer2DModel"]
@@ -74,7 +78,7 @@ if is_torch_available():
     _import_structure["unets.uvit_2d"] = ["UVit2DModel"]
 
 if is_flax_available():
-    _import_structure["controlnet_flax"] = ["FlaxControlNetModel"]
+    _import_structure["controlnets.controlnet_flax"] = ["FlaxControlNetModel"]
     _import_structure["unets.unet_2d_condition_flax"] = ["FlaxUNet2DConditionModel"]
     _import_structure["vae_flax"] = ["FlaxAutoencoderKL"]
 
@@ -94,12 +98,19 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             ConsistencyDecoderVAE,
             VQModel,
         )
-        from .controlnet import ControlNetModel
-        from .controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel
-        from .controlnet_hunyuan import HunyuanDiT2DControlNetModel, HunyuanDiT2DMultiControlNetModel
-        from .controlnet_sd3 import SD3ControlNetModel, SD3MultiControlNetModel
-        from .controlnet_sparsectrl import SparseControlNetModel
-        from .controlnet_xs import ControlNetXSAdapter, UNetControlNetXSModel
+        from .controlnets import (
+            ControlNetModel,
+            ControlNetXSAdapter,
+            FluxControlNetModel,
+            FluxMultiControlNetModel,
+            HunyuanDiT2DControlNetModel,
+            HunyuanDiT2DMultiControlNetModel,
+            MultiControlNetModel,
+            SD3ControlNetModel,
+            SD3MultiControlNetModel,
+            SparseControlNetModel,
+            UNetControlNetXSModel,
+        )
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin
         from .transformers import (
@@ -137,7 +148,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
 
     if is_flax_available():
-        from .controlnet_flax import FlaxControlNetModel
+        from .controlnets import FlaxControlNetModel
         from .unets import FlaxUNet2DConditionModel
         from .vae_flax import FlaxAutoencoderKL
 

src/diffusers/models/controlnet_sd3.py

@@ -13,410 +13,29 @@
 # 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, FusedJointAttnProcessor2_0
-from ..models.modeling_outputs import Transformer2DModelOutput
-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 ..utils import deprecate, logging
+from .controlnets.controlnet_sd3 import SD3ControlNetModel, SD3ControlNetOutput, SD3MultiControlNetModel
 
 
 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,
-        extra_conditioning_channels: int = 0,
-    ):
-        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.config.attention_head_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 + extra_conditioning_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()
-
-            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.transformers.transformer_sd3.SD3Transformer2DModel.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)
-
-        self.set_attn_processor(FusedJointAttnProcessor2_0())
-
-    # 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=12, num_extra_conditioning_channels=1, load_weights_from_transformer=True
-    ):
-        config = transformer.config
-        config["num_layers"] = num_layers or config.num_layers
-        config["extra_conditioning_channels"] = num_extra_conditioning_channels
-        controlnet = cls(**config)
-
-        if load_weights_from_transformer:
-            controlnet.pos_embed.load_state_dict(transformer.pos_embed.state_dict())
-            controlnet.time_text_embed.load_state_dict(transformer.time_text_embed.state_dict())
-            controlnet.context_embedder.load_state_dict(transformer.context_embedder.state_dict())
-            controlnet.transformer_blocks.load_state_dict(transformer.transformer_blocks.state_dict(), strict=False)
-
-            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."
-                )
-
-        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 {}
-                encoder_hidden_states, 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.
-    """
+class SD3ControlNetOutput(SD3ControlNetOutput):
+    def __init__(self, *args, **kwargs):
+        deprecation_message = "Importing `SD3ControlNetOutput` from `diffusers.models.controlnet_sd3` is deprecated and this will be removed in a future version. Please use `from diffusers.models.controlnets.controlnet_sd3 import SD3ControlNetOutput`, instead."
+        deprecate("SD3ControlNetOutput", "0.34", deprecation_message)
+        super().__init__(*args, **kwargs)
 
-    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,
-            )
+class SD3ControlNetModel(SD3ControlNetModel):
+    def __init__(self, *args, **kwargs):
+        deprecation_message = "Importing `SD3ControlNetModel` from `diffusers.models.controlnet_sd3` is deprecated and this will be removed in a future version. Please use `from diffusers.models.controlnets.controlnet_sd3 import SD3ControlNetModel`, instead."
+        deprecate("SD3ControlNetModel", "0.34", deprecation_message)
+        super().__init__(*args, **kwargs)
 
-            # 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
+class SD3MultiControlNetModel(SD3MultiControlNetModel):
+    def __init__(self, *args, **kwargs):
+        deprecation_message = "Importing `SD3MultiControlNetModel` from `diffusers.models.controlnet_sd3` is deprecated and this will be removed in a future version. Please use `from diffusers.models.controlnets.controlnet_sd3 import SD3MultiControlNetModel`, instead."
+        deprecate("SD3MultiControlNetModel", "0.34", deprecation_message)
+        super().__init__(*args, **kwargs)

src/diffusers/models/controlnets/__init__.py

+from ...utils import is_flax_available, is_torch_available
+
+
+if is_torch_available():
+    from .controlnet import ControlNetModel, ControlNetOutput
+    from .controlnet_flux import FluxControlNetModel, FluxControlNetOutput, FluxMultiControlNetModel
+    from .controlnet_hunyuan import (
+        HunyuanControlNetOutput,
+        HunyuanDiT2DControlNetModel,
+        HunyuanDiT2DMultiControlNetModel,
+    )
+    from .controlnet_sd3 import SD3ControlNetModel, SD3ControlNetOutput, SD3MultiControlNetModel
+    from .controlnet_sparsectrl import (
+        SparseControlNetConditioningEmbedding,
+        SparseControlNetModel,
+        SparseControlNetOutput,
+    )
+    from .controlnet_xs import ControlNetXSAdapter, ControlNetXSOutput, UNetControlNetXSModel
+    from .multicontrolnet import MultiControlNetModel
+
+if is_flax_available():
+    from .controlnet_flax import FlaxControlNetModel

src/diffusers/models/controlnet_flax.py → src/diffusers/models/controlnets/controlnet_flax.py

@@ -19,11 +19,11 @@ import jax
 import jax.numpy as jnp
 from flax.core.frozen_dict import FrozenDict
 
-from ..configuration_utils import ConfigMixin, flax_register_to_config
-from ..utils import BaseOutput
-from .embeddings_flax import FlaxTimestepEmbedding, FlaxTimesteps
-from .modeling_flax_utils import FlaxModelMixin
-from .unets.unet_2d_blocks_flax import (
+from ...configuration_utils import ConfigMixin, flax_register_to_config
+from ...utils import BaseOutput
+from ..embeddings_flax import FlaxTimestepEmbedding, FlaxTimesteps
+from ..modeling_flax_utils import FlaxModelMixin
+from ..unets.unet_2d_blocks_flax import (
     FlaxCrossAttnDownBlock2D,
     FlaxDownBlock2D,
     FlaxUNetMidBlock2DCrossAttn,

src/diffusers/models/controlnet_hunyuan.py → src/diffusers/models/controlnets/controlnet_hunyuan.py

@@ -17,17 +17,17 @@ from typing import Dict, Optional, Union
 import torch
 from torch import nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import logging
-from .attention_processor import AttentionProcessor
-from .controlnet import BaseOutput, Tuple, zero_module
-from .embeddings import (
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import logging
+from ..attention_processor import AttentionProcessor
+from ..embeddings import (
     HunyuanCombinedTimestepTextSizeStyleEmbedding,
     PatchEmbed,
     PixArtAlphaTextProjection,
 )
-from .modeling_utils import ModelMixin
-from .transformers.hunyuan_transformer_2d import HunyuanDiTBlock
+from ..modeling_utils import ModelMixin
+from ..transformers.hunyuan_transformer_2d import HunyuanDiTBlock
+from .controlnet import BaseOutput, Tuple, zero_module
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

src/diffusers/models/controlnets/controlnet_sd3.py

+# 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 ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from ..attention import JointTransformerBlock
+from ..attention_processor import Attention, AttentionProcessor, FusedJointAttnProcessor2_0
+from ..embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from .controlnet import BaseOutput, zero_module
+
+
+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,
+        extra_conditioning_channels: int = 0,
+    ):
+        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.config.attention_head_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 + extra_conditioning_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()
+
+            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.transformers.transformer_sd3.SD3Transformer2DModel.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)
+
+        self.set_attn_processor(FusedJointAttnProcessor2_0())
+
+    # 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=12, num_extra_conditioning_channels=1, load_weights_from_transformer=True
+    ):
+        config = transformer.config
+        config["num_layers"] = num_layers or config.num_layers
+        config["extra_conditioning_channels"] = num_extra_conditioning_channels
+        controlnet = cls(**config)
+
+        if load_weights_from_transformer:
+            controlnet.pos_embed.load_state_dict(transformer.pos_embed.state_dict())
+            controlnet.time_text_embed.load_state_dict(transformer.time_text_embed.state_dict())
+            controlnet.context_embedder.load_state_dict(transformer.context_embedder.state_dict())
+            controlnet.transformer_blocks.load_state_dict(transformer.transformer_blocks.state_dict(), strict=False)
+
+            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."
+                )
+
+        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 {}
+                encoder_hidden_states, 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

src/diffusers/models/controlnet_xs.py → src/diffusers/models/controlnets/controlnet_xs.py

@@ -19,10 +19,10 @@ import torch
 import torch.utils.checkpoint
 from torch import Tensor, nn
 
-from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput, is_torch_version, logging
-from ..utils.torch_utils import apply_freeu
-from .attention_processor import (
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput, is_torch_version, logging
+from ...utils.torch_utils import apply_freeu
+from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
     Attention,
@@ -31,10 +31,9 @@ from .attention_processor import (
     AttnProcessor,
     FusedAttnProcessor2_0,
 )
-from .controlnet import ControlNetConditioningEmbedding
-from .embeddings import TimestepEmbedding, Timesteps
-from .modeling_utils import ModelMixin
-from .unets.unet_2d_blocks import (
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from ..unets.unet_2d_blocks import (
     CrossAttnDownBlock2D,
     CrossAttnUpBlock2D,
     Downsample2D,
@@ -43,7 +42,8 @@ from .unets.unet_2d_blocks import (
     UNetMidBlock2DCrossAttn,
     Upsample2D,
 )
-from .unets.unet_2d_condition import UNet2DConditionModel
+from ..unets.unet_2d_condition import UNet2DConditionModel
+from .controlnet import ControlNetConditioningEmbedding
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -1062,7 +1062,8 @@ class UNetControlNetXSModel(ModelMixin, ConfigMixin):
             added_cond_kwargs (`dict`):
                 Additional conditions for the Stable Diffusion XL UNet.
             return_dict (`bool`, defaults to `True`):
-                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+                Whether or not to return a [`~models.controlnets.controlnet.ControlNetOutput`] instead of a plain
+                tuple.
             apply_control (`bool`, defaults to `True`):
                 If `False`, the input is run only through the base model.
 

src/diffusers/models/controlnets/multicontrolnet.py

+import os
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...models.controlnets.controlnet import ControlNetModel, ControlNetOutput
+from ...models.modeling_utils import ModelMixin
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class MultiControlNetModel(ModelMixin):
+    r"""
+    Multiple `ControlNetModel` wrapper class for Multi-ControlNet
+
+    This module is a wrapper for multiple instances of the `ControlNetModel`. The `forward()` API is designed to be
+    compatible with `ControlNetModel`.
+
+    Args:
+        controlnets (`List[ControlNetModel]`):
+            Provides additional conditioning to the unet during the denoising process. You must set multiple
+            `ControlNetModel` as a list.
+    """
+
+    def __init__(self, controlnets: Union[List[ControlNetModel], Tuple[ControlNetModel]]):
+        super().__init__()
+        self.nets = nn.ModuleList(controlnets)
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: List[torch.tensor],
+        conditioning_scale: List[float],
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guess_mode: bool = False,
+        return_dict: bool = True,
+    ) -> Union[ControlNetOutput, Tuple]:
+        for i, (image, scale, controlnet) in enumerate(zip(controlnet_cond, conditioning_scale, self.nets)):
+            down_samples, mid_sample = controlnet(
+                sample=sample,
+                timestep=timestep,
+                encoder_hidden_states=encoder_hidden_states,
+                controlnet_cond=image,
+                conditioning_scale=scale,
+                class_labels=class_labels,
+                timestep_cond=timestep_cond,
+                attention_mask=attention_mask,
+                added_cond_kwargs=added_cond_kwargs,
+                cross_attention_kwargs=cross_attention_kwargs,
+                guess_mode=guess_mode,
+                return_dict=return_dict,
+            )
+
+            # merge samples
+            if i == 0:
+                down_block_res_samples, mid_block_res_sample = down_samples, mid_sample
+            else:
+                down_block_res_samples = [
+                    samples_prev + samples_curr
+                    for samples_prev, samples_curr in zip(down_block_res_samples, down_samples)
+                ]
+                mid_block_res_sample += mid_sample
+
+        return down_block_res_samples, mid_block_res_sample
+
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        is_main_process: bool = True,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        variant: Optional[str] = None,
+    ):
+        """
+        Save a model and its configuration file to a directory, so that it can be re-loaded using the
+        `[`~pipelines.controlnet.MultiControlNetModel.from_pretrained`]` class method.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to which to save. Will be created if it doesn't exist.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful when in distributed training like
+                TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
+                the main process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
+                need to replace `torch.save` by another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+            variant (`str`, *optional*):
+                If specified, weights are saved in the format pytorch_model.<variant>.bin.
+        """
+        for idx, controlnet in enumerate(self.nets):
+            suffix = "" if idx == 0 else f"_{idx}"
+            controlnet.save_pretrained(
+                save_directory + suffix,
+                is_main_process=is_main_process,
+                save_function=save_function,
+                safe_serialization=safe_serialization,
+                variant=variant,
+            )
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_path: Optional[Union[str, os.PathLike]], **kwargs):
+        r"""
+        Instantiate a pretrained MultiControlNet model from multiple pre-trained controlnet models.
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
+        the model, you should first set it back in training mode with `model.train()`.
+
+        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
+        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
+        task.
+
+        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
+        weights are discarded.
+
+        Parameters:
+            pretrained_model_path (`os.PathLike`):
+                A path to a *directory* containing model weights saved using
+                [`~diffusers.pipelines.controlnet.MultiControlNetModel.save_pretrained`], e.g.,
+                `./my_model_directory/controlnet`.
+            torch_dtype (`str` or `torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
+                will be automatically derived from the model's weights.
+            output_loading_info(`bool`, *optional*, defaults to `False`):
+                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+                A map that specifies where each submodule should go. It doesn't need to be refined to each
+                parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
+                same device.
+
+                To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
+                more information about each option see [designing a device
+                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+            max_memory (`Dict`, *optional*):
+                A dictionary device identifier to maximum memory. Will default to the maximum memory available for each
+                GPU and the available CPU RAM if unset.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
+                also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
+                model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
+                setting this argument to `True` will raise an error.
+            variant (`str`, *optional*):
+                If specified load weights from `variant` filename, *e.g.* pytorch_model.<variant>.bin. `variant` is
+                ignored when using `from_flax`.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the `safetensors` weights will be downloaded if they're available **and** if the
+                `safetensors` library is installed. If set to `True`, the model will be forcibly loaded from
+                `safetensors` weights. If set to `False`, loading will *not* use `safetensors`.
+        """
+        idx = 0
+        controlnets = []
+
+        # load controlnet and append to list until no controlnet directory exists anymore
+        # first controlnet has to be saved under `./mydirectory/controlnet` to be compliant with `DiffusionPipeline.from_prertained`
+        # second, third, ... controlnets have to be saved under `./mydirectory/controlnet_1`, `./mydirectory/controlnet_2`, ...
+        model_path_to_load = pretrained_model_path
+        while os.path.isdir(model_path_to_load):
+            controlnet = ControlNetModel.from_pretrained(model_path_to_load, **kwargs)
+            controlnets.append(controlnet)
+
+            idx += 1
+            model_path_to_load = pretrained_model_path + f"_{idx}"
+
+        logger.info(f"{len(controlnets)} controlnets loaded from {pretrained_model_path}.")
+
+        if len(controlnets) == 0:
+            raise ValueError(
+                f"No ControlNets found under {os.path.dirname(pretrained_model_path)}. Expected at least {pretrained_model_path + '_0'}."
+            )
+
+        return cls(controlnets)

src/diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py

@@ -24,7 +24,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPV
 from ...image_processor import PipelineImageInput, VaeImageProcessor
 from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
-from ...models.controlnet_sparsectrl import SparseControlNetModel
+from ...models.controlnets.controlnet_sparsectrl import SparseControlNetModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...models.unets.unet_motion_model import MotionAdapter
 from ...schedulers import KarrasDiffusionSchedulers