diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index ab733054fbd3..27e9fe5e191b 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -400,6 +400,8 @@ title: DiT - local: api/pipelines/flux title: Flux + - local: api/pipelines/control_flux_inpaint + title: FluxControlInpaint - local: api/pipelines/hunyuandit title: Hunyuan-DiT - local: api/pipelines/hunyuan_video diff --git a/docs/source/en/api/pipelines/control_flux_inpaint.md b/docs/source/en/api/pipelines/control_flux_inpaint.md new file mode 100644 index 000000000000..0cf4f4b4225e --- /dev/null +++ b/docs/source/en/api/pipelines/control_flux_inpaint.md @@ -0,0 +1,89 @@ + + +# FluxControlInpaint + +FluxControlInpaintPipeline is an implementation of Inpainting for Flux.1 Depth/Canny models. It is a pipeline that allows you to inpaint images using the Flux.1 Depth/Canny models. The pipeline takes an image and a mask as input and returns the inpainted image. + +FLUX.1 Depth and Canny [dev] is a 12 billion parameter rectified flow transformer capable of generating an image based on a text description while following the structure of a given input image. **This is not a ControlNet model**. + +| Control type | Developer | Link | +| -------- | ---------- | ---- | +| Depth | [Black Forest Labs](https://huggingface.co/black-forest-labs) | [Link](https://huggingface.co/black-forest-labs/FLUX.1-Depth-dev) | +| Canny | [Black Forest Labs](https://huggingface.co/black-forest-labs) | [Link](https://huggingface.co/black-forest-labs/FLUX.1-Canny-dev) | + + + + +Flux can be quite expensive to run on consumer hardware devices. However, you can perform a suite of optimizations to run it faster and in a more memory-friendly manner. Check out [this section](https://huggingface.co/blog/sd3#memory-optimizations-for-sd3) for more details. Additionally, Flux can benefit from quantization for memory efficiency with a trade-off in inference latency. Refer to [this blog post](https://huggingface.co/blog/quanto-diffusers) to learn more. For an exhaustive list of resources, check out [this gist](https://gist.github.com/sayakpaul/b664605caf0aa3bf8585ab109dd5ac9c). + + + +```python +import torch +from diffusers import FluxControlInpaintPipeline +from diffusers.models.transformers import FluxTransformer2DModel +from transformers import T5EncoderModel +from diffusers.utils import load_image, make_image_grid +from image_gen_aux import DepthPreprocessor # https://github.com/huggingface/image_gen_aux +from PIL import Image +import numpy as np + +pipe = FluxControlInpaintPipeline.from_pretrained( + "black-forest-labs/FLUX.1-Depth-dev", + torch_dtype=torch.bfloat16, +) +# use following lines if you have GPU constraints +# --------------------------------------------------------------- +transformer = FluxTransformer2DModel.from_pretrained( + "sayakpaul/FLUX.1-Depth-dev-nf4", subfolder="transformer", torch_dtype=torch.bfloat16 +) +text_encoder_2 = T5EncoderModel.from_pretrained( + "sayakpaul/FLUX.1-Depth-dev-nf4", subfolder="text_encoder_2", torch_dtype=torch.bfloat16 +) +pipe.transformer = transformer +pipe.text_encoder_2 = text_encoder_2 +pipe.enable_model_cpu_offload() +# --------------------------------------------------------------- +pipe.to("cuda") + +prompt = "a blue robot singing opera with human-like expressions" +image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png") + +head_mask = np.zeros_like(image) +head_mask[65:580,300:642] = 255 +mask_image = Image.fromarray(head_mask) + +processor = DepthPreprocessor.from_pretrained("LiheYoung/depth-anything-large-hf") +control_image = processor(image)[0].convert("RGB") + +output = pipe( + prompt=prompt, + image=image, + control_image=control_image, + mask_image=mask_image, + num_inference_steps=30, + strength=0.9, + guidance_scale=10.0, + generator=torch.Generator().manual_seed(42), +).images[0] +make_image_grid([image, control_image, mask_image, output.resize(image.size)], rows=1, cols=4).save("output.png") +``` + +## FluxControlInpaintPipeline +[[autodoc]] FluxControlInpaintPipeline + - all + - __call__ + + +## FluxPipelineOutput +[[autodoc]] pipelines.flux.pipeline_output.FluxPipelineOutput \ No newline at end of file diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index e2351a0c53b8..91b297f8c007 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -277,6 +277,7 @@ "CogView3PlusPipeline", "CycleDiffusionPipeline", "FluxControlImg2ImgPipeline", + "FluxControlInpaintPipeline", "FluxControlNetImg2ImgPipeline", "FluxControlNetInpaintPipeline", "FluxControlNetPipeline", @@ -765,6 +766,7 @@ CogView3PlusPipeline, CycleDiffusionPipeline, FluxControlImg2ImgPipeline, + FluxControlInpaintPipeline, FluxControlNetImg2ImgPipeline, FluxControlNetInpaintPipeline, FluxControlNetPipeline, diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index e7fd7ec78bed..ce291e5ceb45 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -128,6 +128,7 @@ ] _import_structure["flux"] = [ "FluxControlPipeline", + "FluxControlInpaintPipeline", "FluxControlImg2ImgPipeline", "FluxControlNetPipeline", "FluxControlNetImg2ImgPipeline", @@ -539,6 +540,7 @@ ) from .flux import ( FluxControlImg2ImgPipeline, + FluxControlInpaintPipeline, FluxControlNetImg2ImgPipeline, FluxControlNetInpaintPipeline, FluxControlNetPipeline, diff --git a/src/diffusers/pipelines/flux/__init__.py b/src/diffusers/pipelines/flux/__init__.py index 3570368a5ca1..72e1b578f2ca 100644 --- a/src/diffusers/pipelines/flux/__init__.py +++ b/src/diffusers/pipelines/flux/__init__.py @@ -26,6 +26,7 @@ _import_structure["pipeline_flux"] = ["FluxPipeline"] _import_structure["pipeline_flux_control"] = ["FluxControlPipeline"] _import_structure["pipeline_flux_control_img2img"] = ["FluxControlImg2ImgPipeline"] + _import_structure["pipeline_flux_control_inpaint"] = ["FluxControlInpaintPipeline"] _import_structure["pipeline_flux_controlnet"] = ["FluxControlNetPipeline"] _import_structure["pipeline_flux_controlnet_image_to_image"] = ["FluxControlNetImg2ImgPipeline"] _import_structure["pipeline_flux_controlnet_inpainting"] = ["FluxControlNetInpaintPipeline"] @@ -44,6 +45,7 @@ from .pipeline_flux import FluxPipeline from .pipeline_flux_control import FluxControlPipeline from .pipeline_flux_control_img2img import FluxControlImg2ImgPipeline + from .pipeline_flux_control_inpaint import FluxControlInpaintPipeline from .pipeline_flux_controlnet import FluxControlNetPipeline from .pipeline_flux_controlnet_image_to_image import FluxControlNetImg2ImgPipeline from .pipeline_flux_controlnet_inpainting import FluxControlNetInpaintPipeline diff --git a/src/diffusers/pipelines/flux/pipeline_flux_control_inpaint.py b/src/diffusers/pipelines/flux/pipeline_flux_control_inpaint.py new file mode 100644 index 000000000000..a9ac1c72c6ed --- /dev/null +++ b/src/diffusers/pipelines/flux/pipeline_flux_control_inpaint.py @@ -0,0 +1,1141 @@ +# Copyright 2024 Black Forest Labs 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 numpy as np +import torch +from transformers import ( + CLIPTextModel, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +) +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +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 .pipeline_output import FluxPipelineOutput + + +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 FluxControlInpaintPipeline + from diffusers.models.transformers import FluxTransformer2DModel + from transformers import T5EncoderModel + from diffusers.utils import load_image, make_image_grid + from image_gen_aux import DepthPreprocessor # https://github.com/huggingface/image_gen_aux + from PIL import Image + import numpy as np + + pipe = FluxControlInpaintPipeline.from_pretrained( + "black-forest-labs/FLUX.1-Depth-dev", + torch_dtype=torch.bfloat16, + ) + # use following lines if you have GPU constraints + # --------------------------------------------------------------- + transformer = FluxTransformer2DModel.from_pretrained( + "sayakpaul/FLUX.1-Depth-dev-nf4", subfolder="transformer", torch_dtype=torch.bfloat16 + ) + text_encoder_2 = T5EncoderModel.from_pretrained( + "sayakpaul/FLUX.1-Depth-dev-nf4", subfolder="text_encoder_2", torch_dtype=torch.bfloat16 + ) + pipe.transformer = transformer + pipe.text_encoder_2 = text_encoder_2 + pipe.enable_model_cpu_offload() + # --------------------------------------------------------------- + pipe.to("cuda") + + prompt = "a blue robot singing opera with human-like expressions" + image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png") + + head_mask = np.zeros_like(image) + head_mask[65:580, 300:642] = 255 + mask_image = Image.fromarray(head_mask) + + processor = DepthPreprocessor.from_pretrained("LiheYoung/depth-anything-large-hf") + control_image = processor(image)[0].convert("RGB") + + output = pipe( + prompt=prompt, + image=image, + control_image=control_image, + mask_image=mask_image, + num_inference_steps=30, + strength=0.9, + guidance_scale=10.0, + generator=torch.Generator().manual_seed(42), + ).images[0] + make_image_grid([image, control_image, mask_image, output.resize(image.size)], rows=1, cols=4).save( + "output.png" + ) + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.16, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# 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 FluxControlInpaintPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The Flux pipeline for image inpainting using Flux-dev-Depth/Canny. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + 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 ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + 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 + ) + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.vae.config.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + 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 = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._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 = 512, + 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 isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(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_2.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_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.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.flux.pipeline_flux.FluxPipeline._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, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.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 = self.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 = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + prompt_2: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + max_sequence_length: int = 512, + 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 + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + 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. + 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. + 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, FluxLoraLoaderMixin): + 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_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) 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, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # 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.flux.pipeline_flux_img2img.FluxImg2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=None, + 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 height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + 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 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)}") + + 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 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}") + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + 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() + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + 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." + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + shape = (batch_size, num_channels_latents, height, width) + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is not None: + return latents.to(device=device, dtype=dtype), latent_image_ids + + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + return latents, noise, image_latents, latent_image_ids + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + 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_mask_latents( + self, + image, + mask_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + image = self.image_processor.preprocess(image, height=height, width=width) + mask_image = self.mask_processor.preprocess(mask_image, height=height, width=width) + + masked_image = image * (1 - mask_image) + masked_image = masked_image.to(device=device, dtype=dtype) + + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask_image = torch.nn.functional.interpolate(mask_image, size=(height, width)) + mask_image = mask_image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == num_channels_latents: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask_image.shape[0] < batch_size: + if not batch_size % mask_image.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask_image.shape[0]} mask_image were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask_image = mask_image.repeat(batch_size // mask_image.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask_image = self._pack_latents( + mask_image.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + masked_image_latents = torch.cat((masked_image_latents, mask_image), dim=-1) + + return mask_image, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @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, + image: PipelineImageInput = None, + control_image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: Optional[int] = None, + width: Optional[int] = None, + strength: float = 0.6, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 7.0, + 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, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_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 = 512, + ): + 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 + 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. + 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. + mask_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 mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + mask_image_latent (`torch.Tensor`, `List[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will ge generated by `mask_image`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + strength (`float`, *optional*, defaults to 1.0): + 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. + 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 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. + 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. + 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. + 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.flux.FluxPipelineOutput`] 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 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + height, + width, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=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._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + device = self._execution_device + + # 3. 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 + + # 3. Prepare text embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 3. Preprocess mask and image + num_channels_latents = self.vae.config.latent_channels + if masked_image_latents is not None: + # pre computed masked_image_latents and mask_image + masked_image_latents = masked_image_latents.to(latents.device) + mask = mask_image.to(latents.device) + else: + mask, masked_image_latents = self.prepare_mask_latents( + image, + mask_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 4.Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.base_image_seq_len, + self.scheduler.config.max_image_seq_len, + self.scheduler.config.base_shift, + self.scheduler.config.max_shift, + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 8 + + 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, + ) + + if control_image.ndim == 4: + control_image = self.vae.encode(control_image).latent_dist.sample(generator=generator) + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + height_control_image, width_control_image = control_image.shape[2:] + control_image = self._pack_latents( + control_image, + batch_size * num_images_per_prompt, + num_channels_latents, + height_control_image, + width_control_image, + ) + + latents, noise, image_latents, latent_image_ids = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height_8 = 2 * (int(height) // (self.vae_scale_factor * 2)) + width_8 = 2 * (int(width) // (self.vae_scale_factor * 2)) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + # 6. Denoising loop + 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, control_image], dim=2) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # 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] + + # for 64 channel transformer only. + init_mask = mask + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + image_latents, torch.tensor([noise_timestep]), noise + ) + else: + init_latents_proper = image_latents + init_latents_proper = self._pack_latents( + init_latents_proper, batch_size * num_images_per_prompt, num_channels_latents, height_8, width_8 + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + 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) + + # 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 = self._unpack_latents(latents, height, width, self.vae_scale_factor) + 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 not return_dict: + return (image,) + + return FluxPipelineOutput(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 e148c025d191..9b36be9e0604 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -392,6 +392,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch", "transformers"]) +class FluxControlInpaintPipeline(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 FluxControlNetImg2ImgPipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/tests/pipelines/flux/test_pipeline_flux_control_inpaint.py b/tests/pipelines/flux/test_pipeline_flux_control_inpaint.py new file mode 100644 index 000000000000..c5ff02a525f2 --- /dev/null +++ b/tests/pipelines/flux/test_pipeline_flux_control_inpaint.py @@ -0,0 +1,215 @@ +import unittest + +import numpy as np +import torch +from PIL import Image +from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel + +from diffusers import ( + AutoencoderKL, + FlowMatchEulerDiscreteScheduler, + FluxControlInpaintPipeline, + FluxTransformer2DModel, +) +from diffusers.utils.testing_utils import ( + torch_device, +) + +from ..test_pipelines_common import ( + PipelineTesterMixin, + check_qkv_fusion_matches_attn_procs_length, + check_qkv_fusion_processors_exist, +) + + +class FluxControlInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin): + pipeline_class = FluxControlInpaintPipeline + params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"]) + batch_params = frozenset(["prompt"]) + + # there is no xformers processor for Flux + test_xformers_attention = False + + def get_dummy_components(self): + torch.manual_seed(0) + transformer = FluxTransformer2DModel( + patch_size=1, + in_channels=8, + out_channels=4, + num_layers=1, + num_single_layers=1, + attention_head_dim=16, + num_attention_heads=2, + joint_attention_dim=32, + pooled_projection_dim=32, + axes_dims_rope=[4, 4, 8], + ) + 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 = CLIPTextModel(clip_text_encoder_config) + + torch.manual_seed(0) + text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5") + + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_2 = 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=1, + 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, + "tokenizer": tokenizer, + "tokenizer_2": tokenizer_2, + "transformer": transformer, + "vae": vae, + } + + def get_dummy_inputs(self, device, seed=0): + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device="cpu").manual_seed(seed) + + image = Image.new("RGB", (8, 8), 0) + control_image = Image.new("RGB", (8, 8), 0) + mask_image = Image.new("RGB", (8, 8), 255) + + inputs = { + "prompt": "A painting of a squirrel eating a burger", + "control_image": control_image, + "generator": generator, + "image": image, + "mask_image": mask_image, + "strength": 0.8, + "num_inference_steps": 2, + "guidance_scale": 30.0, + "height": 8, + "width": 8, + "max_sequence_length": 48, + "output_type": "np", + } + return inputs + + # def test_flux_different_prompts(self): + # pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + + # inputs = self.get_dummy_inputs(torch_device) + # output_same_prompt = pipe(**inputs).images[0] + + # inputs = self.get_dummy_inputs(torch_device) + # inputs["prompt_2"] = "a different prompt" + # output_different_prompts = pipe(**inputs).images[0] + + # max_diff = np.abs(output_same_prompt - output_different_prompts).max() + + # # Outputs should be different here + # # For some reasons, they don't show large differences + # assert max_diff > 1e-6 + + def test_flux_prompt_embeds(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + inputs = self.get_dummy_inputs(torch_device) + + output_with_prompt = pipe(**inputs).images[0] + + inputs = self.get_dummy_inputs(torch_device) + prompt = inputs.pop("prompt") + + (prompt_embeds, pooled_prompt_embeds, text_ids) = pipe.encode_prompt( + prompt, + prompt_2=None, + device=torch_device, + max_sequence_length=inputs["max_sequence_length"], + ) + output_with_embeds = pipe( + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + **inputs, + ).images[0] + + max_diff = np.abs(output_with_prompt - output_with_embeds).max() + assert max_diff < 1e-4 + + def test_fused_qkv_projections(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe = pipe.to(device) + pipe.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + original_image_slice = image[0, -3:, -3:, -1] + + # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added + # to the pipeline level. + pipe.transformer.fuse_qkv_projections() + assert check_qkv_fusion_processors_exist( + pipe.transformer + ), "Something wrong with the fused attention processors. Expected all the attention processors to be fused." + assert check_qkv_fusion_matches_attn_procs_length( + pipe.transformer, pipe.transformer.original_attn_processors + ), "Something wrong with the attention processors concerning the fused QKV projections." + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + image_slice_fused = image[0, -3:, -3:, -1] + + pipe.transformer.unfuse_qkv_projections() + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + image_slice_disabled = image[0, -3:, -3:, -1] + + assert np.allclose( + original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3 + ), "Fusion of QKV projections shouldn't affect the outputs." + assert np.allclose( + image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3 + ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled." + assert np.allclose( + original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2 + ), "Original outputs should match when fused QKV projections are disabled." + + def test_flux_image_output_shape(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + inputs = self.get_dummy_inputs(torch_device) + + height_width_pairs = [(32, 32), (72, 57)] + for height, width in height_width_pairs: + expected_height = height - height % (pipe.vae_scale_factor * 2) + expected_width = width - width % (pipe.vae_scale_factor * 2) + + inputs.update({"height": height, "width": width}) + image = pipe(**inputs).images[0] + output_height, output_width, _ = image.shape + assert (output_height, output_width) == (expected_height, expected_width)