diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index d39b5a52d2fe..3c85720d8c06 100644 --- a/docs/source/en/_toctree.yml +++ b/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 diff --git a/docs/source/en/api/models/controlnet_sana.md b/docs/source/en/api/models/controlnet_sana.md new file mode 100644 index 000000000000..f0426308f7e0 --- /dev/null +++ b/docs/source/en/api/models/controlnet_sana.md @@ -0,0 +1,29 @@ + + +# 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 + diff --git a/docs/source/en/api/pipelines/controlnet_sana.md b/docs/source/en/api/pipelines/controlnet_sana.md new file mode 100644 index 000000000000..fa04591532c9 --- /dev/null +++ b/docs/source/en/api/pipelines/controlnet_sana.md @@ -0,0 +1,36 @@ + + +# ControlNet + +
+ LoRA +
+ +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 \ No newline at end of file diff --git a/scripts/convert_sana_controlnet_to_diffusers.py b/scripts/convert_sana_controlnet_to_diffusers.py new file mode 100644 index 000000000000..bc1eb327880c --- /dev/null +++ b/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) diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index 3c7e8654d223..e2ce570d5e30 100644 --- a/src/diffusers/__init__.py +++ b/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, diff --git a/src/diffusers/models/__init__.py b/src/diffusers/models/__init__.py index 99a2f871c837..1a11272fa4ba 100755 --- a/src/diffusers/models/__init__.py +++ b/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, diff --git a/src/diffusers/models/controlnets/__init__.py b/src/diffusers/models/controlnets/__init__.py index 1dd92e51a44c..90ef438d2533 100644 --- a/src/diffusers/models/controlnets/__init__.py +++ b/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, diff --git a/src/diffusers/models/controlnets/controlnet_sana.py b/src/diffusers/models/controlnets/controlnet_sana.py new file mode 100644 index 000000000000..7f9d6d9849c1 --- /dev/null +++ b/src/diffusers/models/controlnets/controlnet_sana.py @@ -0,0 +1,290 @@ +# 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. + +from dataclasses import dataclass +from typing import Any, Dict, Optional, Tuple, Union + +import torch +from torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...loaders import PeftAdapterMixin +from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers +from ..attention_processor import AttentionProcessor +from ..embeddings import PatchEmbed, PixArtAlphaTextProjection +from ..modeling_outputs import Transformer2DModelOutput +from ..modeling_utils import ModelMixin +from ..normalization import AdaLayerNormSingle, RMSNorm +from ..transformers.sana_transformer import SanaTransformerBlock +from .controlnet import zero_module + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class SanaControlNetOutput(BaseOutput): + controlnet_block_samples: Tuple[torch.Tensor] + + +class SanaControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin): + _supports_gradient_checkpointing = True + _no_split_modules = ["SanaTransformerBlock", "PatchEmbed"] + _skip_layerwise_casting_patterns = ["patch_embed", "norm"] + + @register_to_config + def __init__( + self, + in_channels: int = 32, + out_channels: Optional[int] = 32, + num_attention_heads: int = 70, + attention_head_dim: int = 32, + num_layers: int = 7, + num_cross_attention_heads: Optional[int] = 20, + cross_attention_head_dim: Optional[int] = 112, + cross_attention_dim: Optional[int] = 2240, + caption_channels: int = 2304, + mlp_ratio: float = 2.5, + dropout: float = 0.0, + attention_bias: bool = False, + sample_size: int = 32, + patch_size: int = 1, + norm_elementwise_affine: bool = False, + norm_eps: float = 1e-6, + interpolation_scale: Optional[int] = None, + ) -> None: + super().__init__() + + out_channels = out_channels or in_channels + inner_dim = num_attention_heads * attention_head_dim + + # 1. Patch Embedding + self.patch_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + interpolation_scale=interpolation_scale, + pos_embed_type="sincos" if interpolation_scale is not None else None, + ) + + # 2. Additional condition embeddings + self.time_embed = AdaLayerNormSingle(inner_dim) + + self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim) + self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True) + + # 3. Transformer blocks + self.transformer_blocks = nn.ModuleList( + [ + SanaTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + num_cross_attention_heads=num_cross_attention_heads, + cross_attention_head_dim=cross_attention_head_dim, + cross_attention_dim=cross_attention_dim, + attention_bias=attention_bias, + norm_elementwise_affine=norm_elementwise_affine, + norm_eps=norm_eps, + mlp_ratio=mlp_ratio, + ) + for _ in range(num_layers) + ] + ) + + # controlnet_blocks + self.controlnet_blocks = nn.ModuleList([]) + + self.input_block = zero_module(nn.Linear(inner_dim, inner_dim)) + for _ in range(len(self.transformer_blocks)): + controlnet_block = nn.Linear(inner_dim, inner_dim) + controlnet_block = zero_module(controlnet_block) + self.controlnet_blocks.append(controlnet_block) + + self.gradient_checkpointing = False + + @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) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + timestep: torch.LongTensor, + controlnet_cond: torch.Tensor, + conditioning_scale: float = 1.0, + encoder_attention_mask: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + attention_kwargs: Optional[Dict[str, Any]] = None, + return_dict: bool = True, + ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]: + if attention_kwargs is not None: + attention_kwargs = attention_kwargs.copy() + lora_scale = 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 attention_kwargs is not None and attention_kwargs.get("scale", None) is not None: + logger.warning( + "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective." + ) + + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. + # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. + # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None and attention_mask.ndim == 2: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: + encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 1. Input + batch_size, num_channels, height, width = hidden_states.shape + p = self.config.patch_size + post_patch_height, post_patch_width = height // p, width // p + + hidden_states = self.patch_embed(hidden_states) + hidden_states = hidden_states + self.input_block(self.patch_embed(controlnet_cond.to(hidden_states.dtype))) + + timestep, embedded_timestep = self.time_embed( + timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype + ) + + encoder_hidden_states = self.caption_projection(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1]) + + encoder_hidden_states = self.caption_norm(encoder_hidden_states) + + # 2. Transformer blocks + block_res_samples = () + if torch.is_grad_enabled() and self.gradient_checkpointing: + for block in self.transformer_blocks: + hidden_states = self._gradient_checkpointing_func( + block, + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + post_patch_height, + post_patch_width, + ) + block_res_samples = block_res_samples + (hidden_states,) + else: + for block in self.transformer_blocks: + hidden_states = block( + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + post_patch_height, + post_patch_width, + ) + block_res_samples = block_res_samples + (hidden_states,) + + # 3. ControlNet blocks + 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,) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + controlnet_block_res_samples = [sample * conditioning_scale for sample in controlnet_block_res_samples] + + if not return_dict: + return (controlnet_block_res_samples,) + + return SanaControlNetOutput(controlnet_block_samples=controlnet_block_res_samples) diff --git a/src/diffusers/models/transformers/sana_transformer.py b/src/diffusers/models/transformers/sana_transformer.py index 48b731406191..52236275dc88 100644 --- a/src/diffusers/models/transformers/sana_transformer.py +++ b/src/diffusers/models/transformers/sana_transformer.py @@ -483,6 +483,7 @@ def forward( encoder_attention_mask: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, attention_kwargs: Optional[Dict[str, Any]] = None, + controlnet_block_samples: Optional[Tuple[torch.Tensor]] = None, return_dict: bool = True, ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]: if attention_kwargs is not None: @@ -546,7 +547,7 @@ def forward( # 2. Transformer blocks if torch.is_grad_enabled() and self.gradient_checkpointing: - for block in self.transformer_blocks: + for index_block, block in enumerate(self.transformer_blocks): hidden_states = self._gradient_checkpointing_func( block, hidden_states, @@ -557,9 +558,11 @@ def forward( post_patch_height, post_patch_width, ) + if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples): + hidden_states = hidden_states + controlnet_block_samples[index_block - 1] else: - for block in self.transformer_blocks: + for index_block, block in enumerate(self.transformer_blocks): hidden_states = block( hidden_states, attention_mask, @@ -569,6 +572,8 @@ def forward( post_patch_height, post_patch_width, ) + if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples): + hidden_states = hidden_states + controlnet_block_samples[index_block - 1] # 3. Normalization hidden_states = self.norm_out(hidden_states, embedded_timestep, self.scale_shift_table) diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index b901d42d9cf7..00aad6bf10d0 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -280,7 +280,7 @@ _import_structure["paint_by_example"] = ["PaintByExamplePipeline"] _import_structure["pia"] = ["PIAPipeline"] _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"] - _import_structure["sana"] = ["SanaPipeline", "SanaSprintPipeline"] + _import_structure["sana"] = ["SanaPipeline", "SanaSprintPipeline", "SanaControlNetPipeline"] _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"] _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"] _import_structure["stable_audio"] = [ @@ -651,7 +651,7 @@ from .paint_by_example import PaintByExamplePipeline from .pia import PIAPipeline from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline - from .sana import SanaPipeline, SanaSprintPipeline + from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintPipeline from .semantic_stable_diffusion import SemanticStableDiffusionPipeline from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline from .stable_audio import StableAudioPipeline, StableAudioProjectionModel diff --git a/src/diffusers/pipelines/sana/__init__.py b/src/diffusers/pipelines/sana/__init__.py index 1393b37e2d3a..5f188ca50815 100644 --- a/src/diffusers/pipelines/sana/__init__.py +++ b/src/diffusers/pipelines/sana/__init__.py @@ -23,6 +23,7 @@ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) else: _import_structure["pipeline_sana"] = ["SanaPipeline"] + _import_structure["pipeline_sana_controlnet"] = ["SanaControlNetPipeline"] _import_structure["pipeline_sana_sprint"] = ["SanaSprintPipeline"] if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: @@ -34,6 +35,7 @@ from ...utils.dummy_torch_and_transformers_objects import * else: from .pipeline_sana import SanaPipeline + from .pipeline_sana_controlnet import SanaControlNetPipeline from .pipeline_sana_sprint import SanaSprintPipeline else: import sys diff --git a/src/diffusers/pipelines/sana/pipeline_sana.py b/src/diffusers/pipelines/sana/pipeline_sana.py index 6093fd836aad..80e0d9bb933f 100644 --- a/src/diffusers/pipelines/sana/pipeline_sana.py +++ b/src/diffusers/pipelines/sana/pipeline_sana.py @@ -354,9 +354,7 @@ def encode_prompt( if device is None: device = self._execution_device - if self.transformer is not None: - dtype = self.transformer.dtype - elif self.text_encoder is not None: + if self.text_encoder is not None: dtype = self.text_encoder.dtype else: dtype = None @@ -928,22 +926,22 @@ def __call__( num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) + transformer_dtype = self.transformer.dtype with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): if self.interrupt: continue latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents - latent_model_input = latent_model_input.to(prompt_embeds.dtype) # broadcast to batch dimension in a way that's compatible with ONNX/Core ML - timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype) + timestep = t.expand(latent_model_input.shape[0]) timestep = timestep * self.transformer.config.timestep_scale # predict noise model_output noise_pred = self.transformer( - latent_model_input, - encoder_hidden_states=prompt_embeds, + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), encoder_attention_mask=prompt_attention_mask, timestep=timestep, return_dict=False, @@ -959,8 +957,6 @@ def __call__( # learned sigma if self.transformer.config.out_channels // 2 == latent_channels: noise_pred = noise_pred.chunk(2, dim=1)[0] - else: - noise_pred = noise_pred # compute previous image: x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] diff --git a/src/diffusers/pipelines/sana/pipeline_sana_controlnet.py b/src/diffusers/pipelines/sana/pipeline_sana_controlnet.py new file mode 100644 index 000000000000..21547d7d4974 --- /dev/null +++ b/src/diffusers/pipelines/sana/pipeline_sana_controlnet.py @@ -0,0 +1,1100 @@ +# Copyright 2024 PixArt-Sigma Authors and 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. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import torch +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, PixArtImageProcessor +from ...loaders import SanaLoraLoaderMixin +from ...models import AutoencoderDC, SanaControlNetModel, SanaTransformer2DModel +from ...schedulers import DPMSolverMultistepScheduler +from ...utils import ( + BACKENDS_MAPPING, + USE_PEFT_BACKEND, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from ..pixart_alpha.pipeline_pixart_alpha import ( + ASPECT_RATIO_512_BIN, + ASPECT_RATIO_1024_BIN, +) +from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN +from .pipeline_output import SanaPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +ASPECT_RATIO_4096_BIN = { + "0.25": [2048.0, 8192.0], + "0.26": [2048.0, 7936.0], + "0.27": [2048.0, 7680.0], + "0.28": [2048.0, 7424.0], + "0.32": [2304.0, 7168.0], + "0.33": [2304.0, 6912.0], + "0.35": [2304.0, 6656.0], + "0.4": [2560.0, 6400.0], + "0.42": [2560.0, 6144.0], + "0.48": [2816.0, 5888.0], + "0.5": [2816.0, 5632.0], + "0.52": [2816.0, 5376.0], + "0.57": [3072.0, 5376.0], + "0.6": [3072.0, 5120.0], + "0.68": [3328.0, 4864.0], + "0.72": [3328.0, 4608.0], + "0.78": [3584.0, 4608.0], + "0.82": [3584.0, 4352.0], + "0.88": [3840.0, 4352.0], + "0.94": [3840.0, 4096.0], + "1.0": [4096.0, 4096.0], + "1.07": [4096.0, 3840.0], + "1.13": [4352.0, 3840.0], + "1.21": [4352.0, 3584.0], + "1.29": [4608.0, 3584.0], + "1.38": [4608.0, 3328.0], + "1.46": [4864.0, 3328.0], + "1.67": [5120.0, 3072.0], + "1.75": [5376.0, 3072.0], + "2.0": [5632.0, 2816.0], + "2.09": [5888.0, 2816.0], + "2.4": [6144.0, 2560.0], + "2.5": [6400.0, 2560.0], + "2.89": [6656.0, 2304.0], + "3.0": [6912.0, 2304.0], + "3.11": [7168.0, 2304.0], + "3.62": [7424.0, 2048.0], + "3.75": [7680.0, 2048.0], + "3.88": [7936.0, 2048.0], + "4.0": [8192.0, 2048.0], +} + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaControlNetPipeline + >>> from diffusers.utils import load_image + + >>> pipe = SanaControlNetPipeline.from_pretrained( + ... "ishan24/Sana_600M_1024px_ControlNetPlus_diffusers", + ... variant="fp16", + ... torch_dtype={"default": torch.bfloat16, "controlnet": torch.float16, "transformer": torch.float16}, + ... device_map="balanced", + ... ) + >>> cond_image = load_image( + ... "https://huggingface.co/ishan24/Sana_600M_1024px_ControlNet_diffusers/resolve/main/hed_example.png" + ... ) + >>> prompt = 'a cat with a neon sign that says "Sana"' + >>> image = pipe( + ... prompt, + ... control_image=cond_image, + ... ).images[0] + >>> image.save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class SanaControlNetPipeline(DiffusionPipeline, SanaLoraLoaderMixin): + r""" + Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629). + """ + + # fmt: off + bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->controlnet->transformer->vae" + _callback_tensor_inputs = ["latents", "control_image", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: Union[GemmaTokenizer, GemmaTokenizerFast], + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC, + transformer: SanaTransformer2DModel, + controlnet: SanaControlNetModel, + scheduler: DPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + vae=vae, + transformer=transformer, + controlnet=controlnet, + scheduler=scheduler, + ) + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + if hasattr(self, "vae") and self.vae is not None + else 32 + ) + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: Union[str, List[str]], + device: torch.device, + dtype: torch.dtype, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: Optional[List[str]] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + prompt = [prompt] if isinstance(prompt, str) else prompt + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_sequence_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_sequence_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + prompt_attention_mask: Optional[torch.Tensor] = None, + negative_prompt_attention_mask: Optional[torch.Tensor] = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: Optional[List[str]] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `List[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + prompt_embeds = prompt_embeds[:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=False, + ) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + if self.text_encoder is not None: + if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip adresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + negative_prompt: str = "", + num_inference_steps: int = 20, + timesteps: List[int] = None, + sigmas: List[float] = None, + guidance_scale: float = 4.5, + control_image: PipelineImageInput = None, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + num_images_per_prompt: Optional[int] = 1, + height: int = 1024, + width: int = 1024, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + prompt_attention_mask: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_attention_mask: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: List[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + ) -> Union[SanaPipelineOutput, Tuple]: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,: + `List[List[torch.Tensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set + the corresponding scale as a list. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + attention_kwargs: + 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). + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `300`): + Maximum sequence length to use with the `prompt`. + complex_human_instruction (`List[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + + Examples: + + Returns: + [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + if use_resolution_binning: + if self.transformer.config.sample_size == 128: + aspect_ratio_bin = ASPECT_RATIO_4096_BIN + elif self.transformer.config.sample_size == 64: + aspect_ratio_bin = ASPECT_RATIO_2048_BIN + elif self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + elif self.transformer.config.sample_size == 16: + aspect_ratio_bin = ASPECT_RATIO_512_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + callback_on_step_end_tensor_inputs, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + lora_scale=lora_scale, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare control image + if isinstance(self.controlnet, SanaControlNetModel): + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + height, width = control_image.shape[-2:] + + control_image = self.vae.encode(control_image).latent + control_image = control_image * self.vae.config.scaling_factor + else: + raise ValueError("`controlnet` must be of type `SanaControlNetModel`.") + + # 5. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 6. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + controlnet_dtype = self.controlnet.dtype + transformer_dtype = self.transformer.dtype + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + # controlnet(s) inference + controlnet_block_samples = self.controlnet( + latent_model_input.to(dtype=controlnet_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=controlnet_dtype), + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + controlnet_cond=control_image, + conditioning_scale=controlnet_conditioning_scale, + )[0] + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + controlnet_block_samples=tuple(t.to(dtype=transformer_dtype) for t in controlnet_block_samples), + )[0] + noise_pred = noise_pred.float() + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + + # compute previous image: x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) + try: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + except torch.cuda.OutOfMemoryError as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return SanaPipelineOutput(images=image) diff --git a/src/diffusers/pipelines/sana/pipeline_sana_sprint.py b/src/diffusers/pipelines/sana/pipeline_sana_sprint.py index 9b3acbb1cb22..30cc8d5f32d0 100644 --- a/src/diffusers/pipelines/sana/pipeline_sana_sprint.py +++ b/src/diffusers/pipelines/sana/pipeline_sana_sprint.py @@ -295,9 +295,7 @@ def encode_prompt( if device is None: device = self._execution_device - if self.transformer is not None: - dtype = self.transformer.dtype - elif self.text_encoder is not None: + if self.text_encoder is not None: dtype = self.text_encoder.dtype else: dtype = None @@ -806,13 +804,14 @@ def __call__( num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) + transformer_dtype = self.transformer.dtype with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): if self.interrupt: continue # broadcast to batch dimension in a way that's compatible with ONNX/Core ML - timestep = t.expand(latents.shape[0]).to(prompt_embeds.dtype) + timestep = t.expand(latents.shape[0]) latents_model_input = latents / self.scheduler.config.sigma_data scm_timestep = torch.sin(timestep) / (torch.cos(timestep) + torch.sin(timestep)) @@ -821,12 +820,11 @@ def __call__( latent_model_input = latents_model_input * torch.sqrt( scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2 ) - latent_model_input = latent_model_input.to(prompt_embeds.dtype) # predict noise model_output noise_pred = self.transformer( - latent_model_input, - encoder_hidden_states=prompt_embeds, + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), encoder_attention_mask=prompt_attention_mask, guidance=guidance, timestep=scm_timestep, diff --git a/src/diffusers/utils/dummy_pt_objects.py b/src/diffusers/utils/dummy_pt_objects.py index dd9117ddca18..5dd488a6cf54 100644 --- a/src/diffusers/utils/dummy_pt_objects.py +++ b/src/diffusers/utils/dummy_pt_objects.py @@ -775,6 +775,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch"]) +class SanaControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + class SanaTransformer2DModel(metaclass=DummyObject): _backends = ["torch"] diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py index b28fba948149..2d424af7749c 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -1502,6 +1502,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch", "transformers"]) +class SanaControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + class SanaPAGPipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/tests/pipelines/sana/test_sana_controlnet.py b/tests/pipelines/sana/test_sana_controlnet.py new file mode 100644 index 000000000000..69a54bc21691 --- /dev/null +++ b/tests/pipelines/sana/test_sana_controlnet.py @@ -0,0 +1,327 @@ +# Copyright 2024 The HuggingFace Team. +# +# 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. + +import inspect +import unittest + +import numpy as np +import torch +from transformers import Gemma2Config, Gemma2Model, GemmaTokenizer + +from diffusers import ( + AutoencoderDC, + FlowMatchEulerDiscreteScheduler, + SanaControlNetModel, + SanaControlNetPipeline, + SanaTransformer2DModel, +) +from diffusers.utils.testing_utils import ( + enable_full_determinism, + torch_device, +) + +from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS +from ..test_pipelines_common import PipelineTesterMixin, to_np + + +enable_full_determinism() + + +class SanaControlNetPipelineFastTests(PipelineTesterMixin, unittest.TestCase): + pipeline_class = SanaControlNetPipeline + params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"} + batch_params = TEXT_TO_IMAGE_BATCH_PARAMS + image_params = TEXT_TO_IMAGE_IMAGE_PARAMS + image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS + required_optional_params = frozenset( + [ + "num_inference_steps", + "generator", + "latents", + "return_dict", + "callback_on_step_end", + "callback_on_step_end_tensor_inputs", + ] + ) + test_xformers_attention = False + test_layerwise_casting = True + test_group_offloading = True + + def get_dummy_components(self): + torch.manual_seed(0) + controlnet = SanaControlNetModel( + patch_size=1, + in_channels=4, + out_channels=4, + num_layers=1, + num_attention_heads=2, + attention_head_dim=4, + num_cross_attention_heads=2, + cross_attention_head_dim=4, + cross_attention_dim=8, + caption_channels=8, + sample_size=32, + ) + + torch.manual_seed(0) + transformer = SanaTransformer2DModel( + patch_size=1, + in_channels=4, + out_channels=4, + num_layers=1, + num_attention_heads=2, + attention_head_dim=4, + num_cross_attention_heads=2, + cross_attention_head_dim=4, + cross_attention_dim=8, + caption_channels=8, + sample_size=32, + ) + + torch.manual_seed(0) + vae = AutoencoderDC( + in_channels=3, + latent_channels=4, + attention_head_dim=2, + encoder_block_types=( + "ResBlock", + "EfficientViTBlock", + ), + decoder_block_types=( + "ResBlock", + "EfficientViTBlock", + ), + encoder_block_out_channels=(8, 8), + decoder_block_out_channels=(8, 8), + encoder_qkv_multiscales=((), (5,)), + decoder_qkv_multiscales=((), (5,)), + encoder_layers_per_block=(1, 1), + decoder_layers_per_block=[1, 1], + downsample_block_type="conv", + upsample_block_type="interpolate", + decoder_norm_types="rms_norm", + decoder_act_fns="silu", + scaling_factor=0.41407, + ) + + torch.manual_seed(0) + scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0) + + torch.manual_seed(0) + text_encoder_config = Gemma2Config( + head_dim=16, + hidden_size=8, + initializer_range=0.02, + intermediate_size=64, + max_position_embeddings=8192, + model_type="gemma2", + num_attention_heads=2, + num_hidden_layers=1, + num_key_value_heads=2, + vocab_size=8, + attn_implementation="eager", + ) + text_encoder = Gemma2Model(text_encoder_config) + tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma") + + components = { + "transformer": transformer, + "vae": vae, + "scheduler": scheduler, + "text_encoder": text_encoder, + "tokenizer": tokenizer, + "controlnet": controlnet, + } + return components + + def get_dummy_inputs(self, device, seed=0): + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device=device).manual_seed(seed) + + control_image = torch.randn(1, 3, 32, 32, generator=generator) + inputs = { + "prompt": "", + "negative_prompt": "", + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 6.0, + "height": 32, + "width": 32, + "max_sequence_length": 16, + "output_type": "pt", + "complex_human_instruction": None, + "control_image": control_image, + "controlnet_conditioning_scale": 1.0, + } + return inputs + + def test_inference(self): + device = "cpu" + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe.to(device) + pipe.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs)[0] + generated_image = image[0] + + self.assertEqual(generated_image.shape, (3, 32, 32)) + expected_image = torch.randn(3, 32, 32) + max_diff = np.abs(generated_image - expected_image).max() + self.assertLessEqual(max_diff, 1e10) + + def test_callback_inputs(self): + sig = inspect.signature(self.pipeline_class.__call__) + has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters + has_callback_step_end = "callback_on_step_end" in sig.parameters + + if not (has_callback_tensor_inputs and has_callback_step_end): + return + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe = pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + self.assertTrue( + hasattr(pipe, "_callback_tensor_inputs"), + f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs", + ) + + def callback_inputs_subset(pipe, i, t, callback_kwargs): + # iterate over callback args + for tensor_name, tensor_value in callback_kwargs.items(): + # check that we're only passing in allowed tensor inputs + assert tensor_name in pipe._callback_tensor_inputs + + return callback_kwargs + + def callback_inputs_all(pipe, i, t, callback_kwargs): + for tensor_name in pipe._callback_tensor_inputs: + assert tensor_name in callback_kwargs + + # iterate over callback args + for tensor_name, tensor_value in callback_kwargs.items(): + # check that we're only passing in allowed tensor inputs + assert tensor_name in pipe._callback_tensor_inputs + + return callback_kwargs + + inputs = self.get_dummy_inputs(torch_device) + + # Test passing in a subset + inputs["callback_on_step_end"] = callback_inputs_subset + inputs["callback_on_step_end_tensor_inputs"] = ["latents"] + output = pipe(**inputs)[0] + + # Test passing in a everything + inputs["callback_on_step_end"] = callback_inputs_all + inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs + output = pipe(**inputs)[0] + + def callback_inputs_change_tensor(pipe, i, t, callback_kwargs): + is_last = i == (pipe.num_timesteps - 1) + if is_last: + callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"]) + return callback_kwargs + + inputs["callback_on_step_end"] = callback_inputs_change_tensor + inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs + output = pipe(**inputs)[0] + assert output.abs().sum() < 1e10 + + def test_attention_slicing_forward_pass( + self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3 + ): + if not self.test_attention_slicing: + return + + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + for component in pipe.components.values(): + if hasattr(component, "set_default_attn_processor"): + component.set_default_attn_processor() + pipe.to(torch_device) + pipe.set_progress_bar_config(disable=None) + + generator_device = "cpu" + inputs = self.get_dummy_inputs(generator_device) + output_without_slicing = pipe(**inputs)[0] + + pipe.enable_attention_slicing(slice_size=1) + inputs = self.get_dummy_inputs(generator_device) + output_with_slicing1 = pipe(**inputs)[0] + + pipe.enable_attention_slicing(slice_size=2) + inputs = self.get_dummy_inputs(generator_device) + output_with_slicing2 = pipe(**inputs)[0] + + if test_max_difference: + max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max() + max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max() + self.assertLess( + max(max_diff1, max_diff2), + expected_max_diff, + "Attention slicing should not affect the inference results", + ) + + def test_vae_tiling(self, expected_diff_max: float = 0.2): + generator_device = "cpu" + components = self.get_dummy_components() + + pipe = self.pipeline_class(**components) + pipe.to("cpu") + pipe.set_progress_bar_config(disable=None) + + # Without tiling + inputs = self.get_dummy_inputs(generator_device) + inputs["height"] = inputs["width"] = 128 + output_without_tiling = pipe(**inputs)[0] + + # With tiling + pipe.vae.enable_tiling( + tile_sample_min_height=96, + tile_sample_min_width=96, + tile_sample_stride_height=64, + tile_sample_stride_width=64, + ) + inputs = self.get_dummy_inputs(generator_device) + inputs["height"] = inputs["width"] = 128 + output_with_tiling = pipe(**inputs)[0] + + self.assertLess( + (to_np(output_without_tiling) - to_np(output_with_tiling)).max(), + expected_diff_max, + "VAE tiling should not affect the inference results", + ) + + # TODO(aryan): Create a dummy gemma model with smol vocab size + @unittest.skip( + "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error." + ) + def test_inference_batch_consistent(self): + pass + + @unittest.skip( + "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error." + ) + def test_inference_batch_single_identical(self): + pass + + def test_float16_inference(self): + # Requires higher tolerance as model seems very sensitive to dtype + super().test_float16_inference(expected_max_diff=0.08)