diff --git a/docs/source/en/api/pipelines/pag.md b/docs/source/en/api/pipelines/pag.md index cc6d075f457f..e723761f6fe0 100644 --- a/docs/source/en/api/pipelines/pag.md +++ b/docs/source/en/api/pipelines/pag.md @@ -96,6 +96,10 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial - all - __call__ +## StableDiffusion3PAGImg2ImgPipeline +[[autodoc]] StableDiffusion3PAGImg2ImgPipeline + - all + - __call__ ## PixArtSigmaPAGPipeline [[autodoc]] PixArtSigmaPAGPipeline diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index a4749af5f61b..6f70a8191629 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -339,6 +339,7 @@ "StableDiffusion3Img2ImgPipeline", "StableDiffusion3InpaintPipeline", "StableDiffusion3PAGPipeline", + "StableDiffusion3PAGImg2ImgPipeline", "StableDiffusion3Pipeline", "StableDiffusionAdapterPipeline", "StableDiffusionAttendAndExcitePipeline", @@ -807,6 +808,7 @@ StableDiffusion3ControlNetPipeline, StableDiffusion3Img2ImgPipeline, StableDiffusion3InpaintPipeline, + StableDiffusion3PAGImg2ImgPipeline, StableDiffusion3PAGPipeline, StableDiffusion3Pipeline, StableDiffusionAdapterPipeline, diff --git a/src/diffusers/models/attention_processor.py b/src/diffusers/models/attention_processor.py index ffbf4a0056c6..7351801368dd 100644 --- a/src/diffusers/models/attention_processor.py +++ b/src/diffusers/models/attention_processor.py @@ -1171,6 +1171,7 @@ def __call__( attn: Attention, hidden_states: torch.FloatTensor, encoder_hidden_states: torch.FloatTensor = None, + attention_mask: Optional[torch.FloatTensor] = None, ) -> torch.FloatTensor: residual = hidden_states diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index 5143b1114fd3..6d3a20511696 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -171,6 +171,7 @@ "KolorsPAGPipeline", "HunyuanDiTPAGPipeline", "StableDiffusion3PAGPipeline", + "StableDiffusion3PAGImg2ImgPipeline", "StableDiffusionPAGPipeline", "StableDiffusionPAGImg2ImgPipeline", "StableDiffusionControlNetPAGPipeline", @@ -589,6 +590,7 @@ HunyuanDiTPAGPipeline, KolorsPAGPipeline, PixArtSigmaPAGPipeline, + StableDiffusion3PAGImg2ImgPipeline, StableDiffusion3PAGPipeline, StableDiffusionControlNetPAGInpaintPipeline, StableDiffusionControlNetPAGPipeline, diff --git a/src/diffusers/pipelines/auto_pipeline.py b/src/diffusers/pipelines/auto_pipeline.py index 0214d7dd6f3c..59ed10758a53 100644 --- a/src/diffusers/pipelines/auto_pipeline.py +++ b/src/diffusers/pipelines/auto_pipeline.py @@ -61,6 +61,7 @@ from .pag import ( HunyuanDiTPAGPipeline, PixArtSigmaPAGPipeline, + StableDiffusion3PAGImg2ImgPipeline, StableDiffusion3PAGPipeline, StableDiffusionControlNetPAGInpaintPipeline, StableDiffusionControlNetPAGPipeline, @@ -129,6 +130,7 @@ ("stable-diffusion", StableDiffusionImg2ImgPipeline), ("stable-diffusion-xl", StableDiffusionXLImg2ImgPipeline), ("stable-diffusion-3", StableDiffusion3Img2ImgPipeline), + ("stable-diffusion-3-pag", StableDiffusion3PAGImg2ImgPipeline), ("if", IFImg2ImgPipeline), ("kandinsky", KandinskyImg2ImgCombinedPipeline), ("kandinsky22", KandinskyV22Img2ImgCombinedPipeline), diff --git a/src/diffusers/pipelines/pag/__init__.py b/src/diffusers/pipelines/pag/__init__.py index 6a6723b58ca9..dfd823b0db27 100644 --- a/src/diffusers/pipelines/pag/__init__.py +++ b/src/diffusers/pipelines/pag/__init__.py @@ -31,6 +31,7 @@ _import_structure["pipeline_pag_pixart_sigma"] = ["PixArtSigmaPAGPipeline"] _import_structure["pipeline_pag_sd"] = ["StableDiffusionPAGPipeline"] _import_structure["pipeline_pag_sd_3"] = ["StableDiffusion3PAGPipeline"] + _import_structure["pipeline_pag_sd_3_img2img"] = ["StableDiffusion3PAGImg2ImgPipeline"] _import_structure["pipeline_pag_sd_animatediff"] = ["AnimateDiffPAGPipeline"] _import_structure["pipeline_pag_sd_img2img"] = ["StableDiffusionPAGImg2ImgPipeline"] _import_structure["pipeline_pag_sd_xl"] = ["StableDiffusionXLPAGPipeline"] @@ -54,6 +55,7 @@ from .pipeline_pag_pixart_sigma import PixArtSigmaPAGPipeline from .pipeline_pag_sd import StableDiffusionPAGPipeline from .pipeline_pag_sd_3 import StableDiffusion3PAGPipeline + from .pipeline_pag_sd_3_img2img import StableDiffusion3PAGImg2ImgPipeline from .pipeline_pag_sd_animatediff import AnimateDiffPAGPipeline from .pipeline_pag_sd_img2img import StableDiffusionPAGImg2ImgPipeline from .pipeline_pag_sd_xl import StableDiffusionXLPAGPipeline diff --git a/src/diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py b/src/diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py new file mode 100644 index 000000000000..54e37e0fd286 --- /dev/null +++ b/src/diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py @@ -0,0 +1,1041 @@ +# Copyright 2024 Stability AI 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import PIL.Image +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3LoraLoaderMixin +from ...models.attention_processor import PAGCFGJointAttnProcessor2_0, PAGJointAttnProcessor2_0 +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + 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 ..stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput +from .pag_utils import PAGMixin + + +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 + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusion3PAGImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusion3PAGImg2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", + ... torch_dtype=torch.float16, + ... pag_applied_layers=["blocks.13"], + ... ) + >>> pipe.to("cuda") + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png" + >>> init_image = load_image(url).convert("RGB") + >>> image = pipe(prompt, image=init_image, guidance_scale=5.0, pag_scale=0.7).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# 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 StableDiffusion3PAGImg2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, PAGMixin): + r""" + [PAG pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/pag) for image-to-image generation + using Stable Diffusion 3. + + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + pag_applied_layers: Union[str, List[str]] = "blocks.1", # 1st transformer block + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + self.set_pag_applied_layers( + pag_applied_layers, pag_attn_processors=(PAGCFGJointAttnProcessor2_0(), PAGJointAttnProcessor2_0()) + ) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + self.tokenizer_max_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, 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(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + clip_skip: Optional[int] = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings 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(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + prompt_2: Union[str, List[str]], + prompt_3: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + clip_skip: Optional[int] = None, + max_sequence_length: int = 256, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + 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`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.FloatTensor`, *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.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # 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, SD3LoraLoaderMixin): + 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 self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + strength, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + 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_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} 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)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + 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." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} 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 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_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.prepare_latents + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + if image.shape[1] == self.vae.config.latent_channels: + init_latents = image + + else: + 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." + ) + + elif isinstance(generator, list): + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = (init_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.scale_noise(init_latents, timestep, noise) + latents = init_latents.to(device=device, dtype=dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @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, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_3: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + strength: float = 0.6, + num_inference_steps: int = 50, + timesteps: List[int] = None, + guidance_scale: float = 7.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = 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 = 256, + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + 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. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + 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. + guidance_scale (`float`, *optional*, defaults to 7.0): + 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. + 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`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + 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.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + 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_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + 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 256): Maximum sequence length to use with the `prompt`. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + strength, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + 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.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + pooled_prompt_embeds = self._prepare_perturbed_attention_guidance( + pooled_prompt_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 3. Preprocess image + image = self.image_processor.preprocess(image) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.transformer.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + # 6. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + 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) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_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 / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.transformer.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py index b76ea3824060..4fc7cd6aefff 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -1397,6 +1397,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch", "transformers"]) +class StableDiffusion3PAGImg2ImgPipeline(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 StableDiffusion3PAGPipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/tests/pipelines/pag/test_pag_sd3_img2img.py b/tests/pipelines/pag/test_pag_sd3_img2img.py new file mode 100644 index 000000000000..bffcd254e2c5 --- /dev/null +++ b/tests/pipelines/pag/test_pag_sd3_img2img.py @@ -0,0 +1,276 @@ +import gc +import inspect +import random +import unittest + +import numpy as np +import torch +from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel + +from diffusers import ( + AutoencoderKL, + AutoPipelineForImage2Image, + FlowMatchEulerDiscreteScheduler, + SD3Transformer2DModel, + StableDiffusion3Img2ImgPipeline, + StableDiffusion3PAGImg2ImgPipeline, +) +from diffusers.utils.testing_utils import ( + enable_full_determinism, + floats_tensor, + load_image, + require_torch_gpu, + slow, + torch_device, +) + +from ..pipeline_params import ( + IMAGE_TO_IMAGE_IMAGE_PARAMS, + TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS, + TEXT_GUIDED_IMAGE_VARIATION_PARAMS, + TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS, +) +from ..test_pipelines_common import ( + PipelineTesterMixin, +) + + +enable_full_determinism() + + +class StableDiffusion3PAGImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin): + pipeline_class = StableDiffusion3PAGImg2ImgPipeline + params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS.union({"pag_scale", "pag_adaptive_scale"}) - {"height", "width"} + required_optional_params = PipelineTesterMixin.required_optional_params - {"latents"} + batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS + image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS + image_latens_params = IMAGE_TO_IMAGE_IMAGE_PARAMS + callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS + + test_xformers_attention = False + + def get_dummy_components(self): + torch.manual_seed(0) + transformer = SD3Transformer2DModel( + sample_size=32, + patch_size=1, + in_channels=4, + num_layers=2, + attention_head_dim=8, + num_attention_heads=4, + caption_projection_dim=32, + joint_attention_dim=32, + pooled_projection_dim=64, + out_channels=4, + ) + clip_text_encoder_config = CLIPTextConfig( + bos_token_id=0, + eos_token_id=2, + hidden_size=32, + intermediate_size=37, + layer_norm_eps=1e-05, + num_attention_heads=4, + num_hidden_layers=5, + pad_token_id=1, + vocab_size=1000, + hidden_act="gelu", + projection_dim=32, + ) + + torch.manual_seed(0) + text_encoder = CLIPTextModelWithProjection(clip_text_encoder_config) + + torch.manual_seed(0) + text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config) + + text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5") + + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_3 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5") + + torch.manual_seed(0) + vae = AutoencoderKL( + sample_size=32, + in_channels=3, + out_channels=3, + block_out_channels=(4,), + layers_per_block=1, + latent_channels=4, + norm_num_groups=1, + use_quant_conv=False, + use_post_quant_conv=False, + shift_factor=0.0609, + scaling_factor=1.5035, + ) + + scheduler = FlowMatchEulerDiscreteScheduler() + + return { + "scheduler": scheduler, + "text_encoder": text_encoder, + "text_encoder_2": text_encoder_2, + "text_encoder_3": text_encoder_3, + "tokenizer": tokenizer, + "tokenizer_2": tokenizer_2, + "tokenizer_3": tokenizer_3, + "transformer": transformer, + "vae": vae, + } + + def get_dummy_inputs(self, device, seed=0): + image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device) + image = image / 2 + 0.5 + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device="cpu").manual_seed(seed) + + inputs = { + "prompt": "A painting of a squirrel eating a burger", + "image": image, + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 5.0, + "output_type": "np", + "pag_scale": 0.7, + } + return inputs + + def test_pag_disable_enable(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + components = self.get_dummy_components() + + # base pipeline (expect same output when pag is disabled) + pipe_sd = StableDiffusion3Img2ImgPipeline(**components) + pipe_sd = pipe_sd.to(device) + pipe_sd.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + del inputs["pag_scale"] + assert ( + "pag_scale" not in inspect.signature(pipe_sd.__call__).parameters + ), f"`pag_scale` should not be a call parameter of the base pipeline {pipe_sd.__class__.__name__}." + out = pipe_sd(**inputs).images[0, -3:, -3:, -1] + + components = self.get_dummy_components() + + # pag disabled with pag_scale=0.0 + pipe_pag = self.pipeline_class(**components) + pipe_pag = pipe_pag.to(device) + pipe_pag.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + inputs["pag_scale"] = 0.0 + out_pag_disabled = pipe_pag(**inputs).images[0, -3:, -3:, -1] + + assert np.abs(out.flatten() - out_pag_disabled.flatten()).max() < 1e-3 + + def test_pag_inference(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + components = self.get_dummy_components() + + pipe_pag = self.pipeline_class(**components, pag_applied_layers=["blocks.0"]) + pipe_pag = pipe_pag.to(device) + pipe_pag.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + image = pipe_pag(**inputs).images + image_slice = image[0, -3:, -3:, -1] + + assert image.shape == ( + 1, + 32, + 32, + 3, + ), f"the shape of the output image should be (1, 32, 32, 3) but got {image.shape}" + + expected_slice = np.array( + [0.66063476, 0.44838923, 0.5484299, 0.7242875, 0.5970012, 0.6015729, 0.53080845, 0.52220416, 0.56397927] + ) + max_diff = np.abs(image_slice.flatten() - expected_slice).max() + self.assertLessEqual(max_diff, 1e-3) + + +@slow +@require_torch_gpu +class StableDiffusion3PAGImg2ImgPipelineIntegrationTests(unittest.TestCase): + pipeline_class = StableDiffusion3PAGImg2ImgPipeline + repo_id = "stabilityai/stable-diffusion-3-medium-diffusers" + + def setUp(self): + super().setUp() + gc.collect() + torch.cuda.empty_cache() + + def tearDown(self): + super().tearDown() + gc.collect() + torch.cuda.empty_cache() + + def get_inputs( + self, device, generator_device="cpu", dtype=torch.float32, seed=0, guidance_scale=7.0, pag_scale=0.7 + ): + img_url = ( + "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sdxl-text2img.png" + ) + init_image = load_image(img_url) + + generator = torch.Generator(device=generator_device).manual_seed(seed) + inputs = { + "prompt": "an astronaut in a space suit walking through a jungle", + "generator": generator, + "image": init_image, + "num_inference_steps": 12, + "strength": 0.6, + "guidance_scale": guidance_scale, + "pag_scale": pag_scale, + "output_type": "np", + } + return inputs + + def test_pag_cfg(self): + pipeline = AutoPipelineForImage2Image.from_pretrained( + self.repo_id, enable_pag=True, torch_dtype=torch.float16, pag_applied_layers=["blocks.17"] + ) + pipeline.enable_model_cpu_offload() + pipeline.set_progress_bar_config(disable=None) + + inputs = self.get_inputs(torch_device) + image = pipeline(**inputs).images + image_slice = image[0, -3:, -3:, -1].flatten() + assert image.shape == (1, 1024, 1024, 3) + expected_slice = np.array( + [ + 0.16772461, + 0.17626953, + 0.18432617, + 0.17822266, + 0.18359375, + 0.17626953, + 0.17407227, + 0.17700195, + 0.17822266, + ] + ) + assert ( + np.abs(image_slice.flatten() - expected_slice).max() < 1e-3 + ), f"output is different from expected, {image_slice.flatten()}" + + def test_pag_uncond(self): + pipeline = AutoPipelineForImage2Image.from_pretrained( + self.repo_id, enable_pag=True, torch_dtype=torch.float16, pag_applied_layers=["blocks.(4|17)"] + ) + pipeline.enable_model_cpu_offload() + pipeline.set_progress_bar_config(disable=None) + + inputs = self.get_inputs(torch_device, guidance_scale=0.0, pag_scale=1.8) + image = pipeline(**inputs).images + image_slice = image[0, -3:, -3:, -1].flatten() + assert image.shape == (1, 1024, 1024, 3) + expected_slice = np.array( + [0.1508789, 0.16210938, 0.17138672, 0.16210938, 0.17089844, 0.16137695, 0.16235352, 0.16430664, 0.16455078] + ) + assert ( + np.abs(image_slice.flatten() - expected_slice).max() < 1e-3 + ), f"output is different from expected, {image_slice.flatten()}"