Merge branch 'main' into flux_guidancecontrol_inpaint

Author: sayakpaul

docs/source/en/_toctree.yml

@@ -252,6 +252,8 @@
         title: SD3ControlNetModel
       - local: api/models/controlnet_sparsectrl
         title: SparseControlNetModel
+      - local: api/models/controlnet_union
+        title: ControlNetUnionModel
       title: ControlNets
     - sections:
       - local: api/models/allegro_transformer3d
@@ -368,6 +370,8 @@
       title: ControlNet-XS
     - local: api/pipelines/controlnetxs_sdxl
       title: ControlNet-XS with Stable Diffusion XL
+    - local: api/pipelines/controlnet_union
+      title: ControlNetUnion
     - local: api/pipelines/dance_diffusion
       title: Dance Diffusion
     - local: api/pipelines/ddim

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

@@ -0,0 +1,35 @@
+<!--Copyright 2024 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.
+-->
+
+# ControlNetUnionModel
+
+ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL.
+
+The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation.
+
+*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.*
+
+## Loading
+
+By default the [`ControlNetUnionModel`] should be loaded with [`~ModelMixin.from_pretrained`].
+
+```py
+from diffusers import StableDiffusionXLControlNetUnionPipeline, ControlNetUnionModel
+
+controlnet = ControlNetUnionModel.from_pretrained("xinsir/controlnet-union-sdxl-1.0")
+pipe = StableDiffusionXLControlNetUnionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet)
+```
+
+## ControlNetUnionModel
+
+[[autodoc]] ControlNetUnionModel
+

docs/source/en/api/pipelines/controlnet_union.md

@@ -0,0 +1,35 @@
+<!--Copyright 2024 The HuggingFace 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.
+-->
+
+# ControlNetUnion
+
+ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL.
+
+The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation.
+
+*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.*
+
+
+## StableDiffusionXLControlNetUnionPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLControlNetUnionImg2ImgPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionImg2ImgPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLControlNetUnionInpaintPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionInpaintPipeline
+	- all
+	- __call__

examples/dreambooth/train_dreambooth.py

@@ -1300,16 +1300,17 @@ def compute_text_embeddings(prompt):
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    base_weight = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
 
                     if noise_scheduler.config.prediction_type == "v_prediction":
                         # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
+                        divisor = snr + 1
                     else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
+                        divisor = snr
+
+                    mse_loss_weights = (
+                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / divisor
+                    )
+
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

src/diffusers/__init__.py

@@ -92,6 +92,7 @@
             "CogView3PlusTransformer2DModel",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
+            "ControlNetUnionModel",
             "ControlNetXSAdapter",
             "DiTTransformer2DModel",
             "FluxControlNetModel",
@@ -379,6 +380,9 @@
             "StableDiffusionXLControlNetPAGImg2ImgPipeline",
             "StableDiffusionXLControlNetPAGPipeline",
             "StableDiffusionXLControlNetPipeline",
+            "StableDiffusionXLControlNetUnionImg2ImgPipeline",
+            "StableDiffusionXLControlNetUnionInpaintPipeline",
+            "StableDiffusionXLControlNetUnionPipeline",
             "StableDiffusionXLControlNetXSPipeline",
             "StableDiffusionXLImg2ImgPipeline",
             "StableDiffusionXLInpaintPipeline",
@@ -587,6 +591,7 @@
             CogView3PlusTransformer2DModel,
             ConsistencyDecoderVAE,
             ControlNetModel,
+            ControlNetUnionModel,
             ControlNetXSAdapter,
             DiTTransformer2DModel,
             FluxControlNetModel,
@@ -852,6 +857,9 @@
             StableDiffusionXLControlNetPAGImg2ImgPipeline,
             StableDiffusionXLControlNetPAGPipeline,
             StableDiffusionXLControlNetPipeline,
+            StableDiffusionXLControlNetUnionImg2ImgPipeline,
+            StableDiffusionXLControlNetUnionInpaintPipeline,
+            StableDiffusionXLControlNetUnionPipeline,
             StableDiffusionXLControlNetXSPipeline,
             StableDiffusionXLImg2ImgPipeline,
             StableDiffusionXLInpaintPipeline,

src/diffusers/models/__init__.py

@@ -45,6 +45,7 @@
     ]
     _import_structure["controlnets.controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"]
+    _import_structure["controlnets.controlnet_union"] = ["ControlNetUnionModel"]
     _import_structure["controlnets.controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
     _import_structure["controlnets.multicontrolnet"] = ["MultiControlNetModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
@@ -102,6 +103,7 @@
         )
         from .controlnets import (
             ControlNetModel,
+            ControlNetUnionModel,
             ControlNetXSAdapter,
             FluxControlNetModel,
             FluxMultiControlNetModel,

src/diffusers/models/attention_processor.py

@@ -358,6 +358,14 @@ def set_use_memory_efficient_attention_xformers(
             self.processor,
             (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, IPAdapterXFormersAttnProcessor),
         )
+        is_joint_processor = hasattr(self, "processor") and isinstance(
+            self.processor,
+            (
+                JointAttnProcessor2_0,
+                XFormersJointAttnProcessor,
+            ),
+        )
+
         if use_memory_efficient_attention_xformers:
             if is_added_kv_processor and is_custom_diffusion:
                 raise NotImplementedError(
@@ -420,6 +428,8 @@ def set_use_memory_efficient_attention_xformers(
                     processor.to(
                         device=self.processor.to_k_ip[0].weight.device, dtype=self.processor.to_k_ip[0].weight.dtype
                     )
+            elif is_joint_processor:
+                processor = XFormersJointAttnProcessor(attention_op=attention_op)
             else:
                 processor = XFormersAttnProcessor(attention_op=attention_op)
         else:
@@ -1685,6 +1695,91 @@ def __call__(
         return hidden_states, encoder_hidden_states
 
 
+class XFormersJointAttnProcessor:
+    r"""
+    Processor for implementing memory efficient attention using xFormers.
+
+    Args:
+        attention_op (`Callable`, *optional*, defaults to `None`):
+            The base
+            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
+            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
+            operator.
+    """
+
+    def __init__(self, attention_op: Optional[Callable] = None):
+        self.attention_op = attention_op
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        # `sample` projections.
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # `context` projections.
+        if encoder_hidden_states is not None:
+            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 = attn.head_to_batch_dim(encoder_hidden_states_query_proj).contiguous()
+            encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj).contiguous()
+            encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj).contiguous()
+
+            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)
+
+            query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
+            key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
+            value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)
+
+        hidden_states = xformers.ops.memory_efficient_attention(
+            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
+        )
+        hidden_states = hidden_states.to(query.dtype)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if encoder_hidden_states is not None:
+            # Split the attention outputs.
+            hidden_states, encoder_hidden_states = (
+                hidden_states[:, : residual.shape[1]],
+                hidden_states[:, residual.shape[1] :],
+            )
+            if not attn.context_pre_only:
+                encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if encoder_hidden_states is not None:
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
 class AllegroAttnProcessor2_0:
     r"""
     Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is

src/diffusers/models/controlnets/__init__.py

@@ -15,6 +15,7 @@
         SparseControlNetModel,
         SparseControlNetOutput,
     )
+    from .controlnet_union import ControlNetUnionInput, ControlNetUnionInputProMax, ControlNetUnionModel
     from .controlnet_xs import ControlNetXSAdapter, ControlNetXSOutput, UNetControlNetXSModel
     from .multicontrolnet import MultiControlNetModel