[ControlNet] Adds controlnet for SanaTransformer

docs/source/en/_toctree.yml

@@ -268,16 +268,18 @@
     - sections:
       - local: api/models/controlnet
         title: ControlNetModel
+      - local: api/models/controlnet_union
+        title: ControlNetUnionModel
       - local: api/models/controlnet_flux
         title: FluxControlNetModel
       - local: api/models/controlnet_hunyuandit
         title: HunyuanDiT2DControlNetModel
+      - local: api/models/controlnet_sana
+        title: SanaControlNetModel
       - local: api/models/controlnet_sd3
         title: SD3ControlNetModel
       - local: api/models/controlnet_sparsectrl
         title: SparseControlNetModel
-      - local: api/models/controlnet_union
-        title: ControlNetUnionModel
       title: ControlNets
     - sections:
       - local: api/models/allegro_transformer3d
@@ -420,6 +422,8 @@
       title: ControlNet with Stable Diffusion 3
     - local: api/pipelines/controlnet_sdxl
       title: ControlNet with Stable Diffusion XL
+    - local: api/pipelines/controlnet_sana
+      title: ControlNet-Sana
     - local: api/pipelines/controlnetxs
       title: ControlNet-XS
     - local: api/pipelines/controlnetxs_sdxl

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

@@ -0,0 +1,29 @@
+<!--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.
+-->
+
+# SanaControlNetModel
+
+The ControlNet model was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, Maneesh Agrawala. It provides a greater degree of control over text-to-image generation by conditioning the model on additional inputs such as edge maps, depth maps, segmentation maps, and keypoints for pose detection.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This model was contributed by [ishan24](https://huggingface.co/ishan24). ❤️
+The original codebase can be found at [NVlabs/Sana](https://github.com/NVlabs/Sana), and you can find official ControlNet checkpoints on [Efficient-Large-Model's](https://huggingface.co/Efficient-Large-Model) Hub profile.
+
+## SanaControlNetModel
+[[autodoc]] SanaControlNetModel
+
+## SanaControlNetOutput
+[[autodoc]] models.controlnets.controlnet_sana.SanaControlNetOutput
+

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

@@ -0,0 +1,36 @@
+<!--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.
+-->
+
+# ControlNet
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
+ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
+
+With a ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This pipeline was contributed by [ishan24](https://huggingface.co/ishan24). ❤️
+The original codebase can be found at [NVlabs/Sana](https://github.com/NVlabs/Sana), and you can find official ControlNet checkpoints on [Efficient-Large-Model's](https://huggingface.co/Efficient-Large-Model) Hub profile.
+
+## SanaControlNetPipeline
+[[autodoc]] SanaControlNetPipeline
+	- all
+	- __call__
+
+## SanaPipelineOutput
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput

scripts/convert_sana_controlnet_to_diffusers.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+from contextlib import nullcontext
+
+import torch
+from accelerate import init_empty_weights
+
+from diffusers import (
+    SanaControlNetModel,
+)
+from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+
+def main(args):
+    file_path = args.orig_ckpt_path
+
+    all_state_dict = torch.load(file_path, weights_only=True)
+    state_dict = all_state_dict.pop("state_dict")
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["patch_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["patch_embed.proj.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    # AdaLN-single LN
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["time_embed.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["time_embed.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    # y norm
+    converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
+
+    # Positional embedding interpolation scale.
+    interpolation_scale = {512: None, 1024: None, 2048: 1.0, 4096: 2.0}
+
+    # ControlNet Input Projection.
+    converted_state_dict["input_block.weight"] = state_dict.pop("controlnet.0.before_proj.weight")
+    converted_state_dict["input_block.bias"] = state_dict.pop("controlnet.0.before_proj.bias")
+
+    for depth in range(7):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.scale_shift_table"
+        )
+
+        # Linear Attention is all you need 🤘
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"controlnet.{depth}.copied_block.attn.qkv.weight"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.attn.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.inverted_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.inverted_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.depth_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.depth_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_point.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.point_conv.conv.weight"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(
+            state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.kv_linear.bias"), 2, dim=0
+        )
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.cross_attn.proj.bias"
+        )
+
+        # ControlNet After Projection
+        converted_state_dict[f"controlnet_blocks.{depth}.weight"] = state_dict.pop(
+            f"controlnet.{depth}.after_proj.weight"
+        )
+        converted_state_dict[f"controlnet_blocks.{depth}.bias"] = state_dict.pop(f"controlnet.{depth}.after_proj.bias")
+
+    # ControlNet
+    with CTX():
+        controlnet = SanaControlNetModel(
+            num_attention_heads=model_kwargs[args.model_type]["num_attention_heads"],
+            attention_head_dim=model_kwargs[args.model_type]["attention_head_dim"],
+            num_layers=model_kwargs[args.model_type]["num_layers"],
+            num_cross_attention_heads=model_kwargs[args.model_type]["num_cross_attention_heads"],
+            cross_attention_head_dim=model_kwargs[args.model_type]["cross_attention_head_dim"],
+            cross_attention_dim=model_kwargs[args.model_type]["cross_attention_dim"],
+            caption_channels=2304,
+            sample_size=args.image_size // 32,
+            interpolation_scale=interpolation_scale[args.image_size],
+        )
+
+    if is_accelerate_available():
+        load_model_dict_into_meta(controlnet, converted_state_dict)
+    else:
+        controlnet.load_state_dict(converted_state_dict, strict=True, assign=True)
+
+    num_model_params = sum(p.numel() for p in controlnet.parameters())
+    print(f"Total number of controlnet parameters: {num_model_params}")
+
+    controlnet = controlnet.to(weight_dtype)
+    controlnet.load_state_dict(converted_state_dict, strict=True)
+
+    print(f"Saving Sana ControlNet in Diffusers format in {args.dump_path}.")
+    controlnet.save_pretrained(args.dump_path)
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+VARIANT_MAPPING = {
+    "fp32": None,
+    "fp16": "fp16",
+    "bf16": "bf16",
+}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=1024,
+        type=int,
+        choices=[512, 1024, 2048, 4096],
+        required=False,
+        help="Image size of pretrained model, 512, 1024, 2048 or 4096.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default="SanaMS_1600M_P1_ControlNet_D7",
+        type=str,
+        choices=["SanaMS_1600M_P1_ControlNet_D7", "SanaMS_600M_P1_ControlNet_D7"],
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--dtype", default="fp16", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")
+
+    args = parser.parse_args()
+
+    model_kwargs = {
+        "SanaMS_1600M_P1_ControlNet_D7": {
+            "num_attention_heads": 70,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "num_layers": 7,
+        },
+        "SanaMS_600M_P1_ControlNet_D7": {
+            "num_attention_heads": 36,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 16,
+            "cross_attention_head_dim": 72,
+            "cross_attention_dim": 1152,
+            "num_layers": 7,
+        },
+    }
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    weight_dtype = DTYPE_MAPPING[args.dtype]
+    variant = VARIANT_MAPPING[args.dtype]
+
+    main(args)

src/diffusers/__init__.py

@@ -188,6 +188,7 @@
             "OmniGenTransformer2DModel",
             "PixArtTransformer2DModel",
             "PriorTransformer",
+            "SanaControlNetModel",
             "SanaTransformer2DModel",
             "SD3ControlNetModel",
             "SD3MultiControlNetModel",
@@ -426,6 +427,7 @@
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
             "ReduxImageEncoder",
+            "SanaControlNetPipeline",
             "SanaPAGPipeline",
             "SanaPipeline",
             "SanaSprintPipeline",
@@ -766,6 +768,7 @@
             OmniGenTransformer2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            SanaControlNetModel,
             SanaTransformer2DModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
@@ -983,6 +986,7 @@
             PixArtSigmaPAGPipeline,
             PixArtSigmaPipeline,
             ReduxImageEncoder,
+            SanaControlNetPipeline,
             SanaPAGPipeline,
             SanaPipeline,
             SanaSprintPipeline,

src/diffusers/models/__init__.py

@@ -49,6 +49,7 @@
         "HunyuanDiT2DControlNetModel",
         "HunyuanDiT2DMultiControlNetModel",
     ]
+    _import_structure["controlnets.controlnet_sana"] = ["SanaControlNetModel"]
     _import_structure["controlnets.controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"]
     _import_structure["controlnets.controlnet_union"] = ["ControlNetUnionModel"]
@@ -133,6 +134,7 @@
             HunyuanDiT2DMultiControlNetModel,
             MultiControlNetModel,
             MultiControlNetUnionModel,
+            SanaControlNetModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
             SparseControlNetModel,

src/diffusers/models/controlnets/__init__.py

@@ -9,6 +9,7 @@
         HunyuanDiT2DControlNetModel,
         HunyuanDiT2DMultiControlNetModel,
     )
+    from .controlnet_sana import SanaControlNetModel
     from .controlnet_sd3 import SD3ControlNetModel, SD3ControlNetOutput, SD3MultiControlNetModel
     from .controlnet_sparsectrl import (
         SparseControlNetConditioningEmbedding,