diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index e27c908ece51..81051b9f2542 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -353,6 +353,7 @@ "AuraFlowPipeline", "BlipDiffusionControlNetPipeline", "BlipDiffusionPipeline", + "ChromaImg2ImgPipeline", "ChromaPipeline", "CLIPImageProjection", "CogVideoXFunControlPipeline", @@ -945,6 +946,7 @@ AudioLDM2UNet2DConditionModel, AudioLDMPipeline, AuraFlowPipeline, + ChromaImg2ImgPipeline, ChromaPipeline, CLIPImageProjection, CogVideoXFunControlPipeline, diff --git a/src/diffusers/models/attention_processor.py b/src/diffusers/models/attention_processor.py index 23ae05e2ab96..291e96eb6d6d 100755 --- a/src/diffusers/models/attention_processor.py +++ b/src/diffusers/models/attention_processor.py @@ -2543,7 +2543,9 @@ def __call__( query = apply_rotary_emb(query, image_rotary_emb) key = apply_rotary_emb(key, image_rotary_emb) - hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False) + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) @@ -2776,7 +2778,9 @@ def __call__( query = apply_rotary_emb(query, image_rotary_emb) key = apply_rotary_emb(key, image_rotary_emb) - hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False) + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) diff --git a/src/diffusers/models/transformers/transformer_chroma.py b/src/diffusers/models/transformers/transformer_chroma.py index 2b415cfed2fe..d11f6c2a5e25 100644 --- a/src/diffusers/models/transformers/transformer_chroma.py +++ b/src/diffusers/models/transformers/transformer_chroma.py @@ -250,15 +250,21 @@ def forward( hidden_states: torch.Tensor, temb: torch.Tensor, image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + attention_mask: Optional[torch.Tensor] = None, joint_attention_kwargs: Optional[Dict[str, Any]] = None, ) -> torch.Tensor: residual = hidden_states norm_hidden_states, gate = self.norm(hidden_states, emb=temb) mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states)) joint_attention_kwargs = joint_attention_kwargs or {} + + if attention_mask is not None: + attention_mask = attention_mask[:, None, None, :] * attention_mask[:, None, :, None] + attn_output = self.attn( hidden_states=norm_hidden_states, image_rotary_emb=image_rotary_emb, + attention_mask=attention_mask, **joint_attention_kwargs, ) @@ -312,6 +318,7 @@ def forward( encoder_hidden_states: torch.Tensor, temb: torch.Tensor, image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + attention_mask: Optional[torch.Tensor] = None, joint_attention_kwargs: Optional[Dict[str, Any]] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: temb_img, temb_txt = temb[:, :6], temb[:, 6:] @@ -321,11 +328,15 @@ def forward( encoder_hidden_states, emb=temb_txt ) joint_attention_kwargs = joint_attention_kwargs or {} + if attention_mask is not None: + attention_mask = attention_mask[:, None, None, :] * attention_mask[:, None, :, None] + # Attention. attention_outputs = self.attn( hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states, image_rotary_emb=image_rotary_emb, + attention_mask=attention_mask, **joint_attention_kwargs, ) @@ -570,6 +581,7 @@ def forward( timestep: torch.LongTensor = None, img_ids: torch.Tensor = None, txt_ids: torch.Tensor = None, + attention_mask: torch.Tensor = None, joint_attention_kwargs: Optional[Dict[str, Any]] = None, controlnet_block_samples=None, controlnet_single_block_samples=None, @@ -659,11 +671,7 @@ def forward( ) if torch.is_grad_enabled() and self.gradient_checkpointing: encoder_hidden_states, hidden_states = self._gradient_checkpointing_func( - block, - hidden_states, - encoder_hidden_states, - temb, - image_rotary_emb, + block, hidden_states, encoder_hidden_states, temb, image_rotary_emb, attention_mask ) else: @@ -672,6 +680,7 @@ def forward( encoder_hidden_states=encoder_hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, + attention_mask=attention_mask, joint_attention_kwargs=joint_attention_kwargs, ) @@ -704,6 +713,7 @@ def forward( hidden_states=hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, + attention_mask=attention_mask, joint_attention_kwargs=joint_attention_kwargs, ) diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index 2fca49f51f74..b32d55bd5171 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -148,7 +148,7 @@ "AudioLDM2UNet2DConditionModel", ] _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"] - _import_structure["chroma"] = ["ChromaPipeline"] + _import_structure["chroma"] = ["ChromaPipeline", "ChromaImg2ImgPipeline"] _import_structure["cogvideo"] = [ "CogVideoXPipeline", "CogVideoXImageToVideoPipeline", @@ -537,7 +537,7 @@ ) from .aura_flow import AuraFlowPipeline from .blip_diffusion import BlipDiffusionPipeline - from .chroma import ChromaPipeline + from .chroma import ChromaImg2ImgPipeline, ChromaPipeline from .cogvideo import ( CogVideoXFunControlPipeline, CogVideoXImageToVideoPipeline, diff --git a/src/diffusers/pipelines/chroma/__init__.py b/src/diffusers/pipelines/chroma/__init__.py index 9faa7902a15c..d9238b735c41 100644 --- a/src/diffusers/pipelines/chroma/__init__.py +++ b/src/diffusers/pipelines/chroma/__init__.py @@ -23,6 +23,7 @@ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) else: _import_structure["pipeline_chroma"] = ["ChromaPipeline"] + _import_structure["pipeline_chroma_img2img"] = ["ChromaImg2ImgPipeline"] if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: try: if not (is_transformers_available() and is_torch_available()): @@ -31,6 +32,7 @@ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 else: from .pipeline_chroma import ChromaPipeline + from .pipeline_chroma_img2img import ChromaImg2ImgPipeline else: import sys diff --git a/src/diffusers/pipelines/chroma/pipeline_chroma.py b/src/diffusers/pipelines/chroma/pipeline_chroma.py index c111458d3320..6ce1b7e5f001 100644 --- a/src/diffusers/pipelines/chroma/pipeline_chroma.py +++ b/src/diffusers/pipelines/chroma/pipeline_chroma.py @@ -1,4 +1,4 @@ -# Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved. +# Copyright 2025 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. @@ -52,12 +52,21 @@ >>> import torch >>> from diffusers import ChromaPipeline - >>> pipe = ChromaPipeline.from_single_file( - ... "chroma-unlocked-v35-detail-calibrated.safetensors", torch_dtype=torch.bfloat16 + >>> ckpt_path = "https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v37.safetensors" + >>> transformer = ChromaTransformer2DModel.from_single_file(ckpt_path, torch_dtype=torch.bfloat16) + >>> text_encoder = AutoModel.from_pretrained("black-forest-labs/FLUX.1-schnell", subfolder="text_encoder_2") + >>> tokenizer = AutoTokenizer.from_pretrained("black-forest-labs/FLUX.1-schnell", subfolder="tokenizer_2") + >>> pipe = ChromaImg2ImgPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-schnell", + ... transformer=transformer, + ... text_encoder=text_encoder, + ... tokenizer=tokenizer, + ... torch_dtype=torch.bfloat16, ... ) - >>> pipe.to("cuda") + >>> pipe.enable_model_cpu_offload() >>> prompt = "A cat holding a sign that says hello world" - >>> image = pipe(prompt, num_inference_steps=28, guidance_scale=4.0).images[0] + >>> negative_prompt = "low quality, ugly, unfinished, out of focus, deformed, disfigure, blurry, smudged, restricted palette, flat colors" + >>> image = pipe(prompt, negative_prompt=negative_prompt).images[0] >>> image.save("chroma.png") ``` """ @@ -235,6 +244,7 @@ def _get_t5_prompt_embeds( dtype = self.text_encoder.dtype prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + attention_mask = attention_mask.to(dtype=dtype, device=device) _, seq_len, _ = prompt_embeds.shape @@ -242,7 +252,10 @@ def _get_t5_prompt_embeds( 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 + attention_mask = attention_mask.repeat(1, num_images_per_prompt) + attention_mask = attention_mask.view(batch_size * num_images_per_prompt, seq_len) + + return prompt_embeds, attention_mask def encode_prompt( self, @@ -250,8 +263,10 @@ def encode_prompt( negative_prompt: Union[str, List[str]] = None, device: Optional[torch.device] = None, num_images_per_prompt: int = 1, - prompt_embeds: Optional[torch.FloatTensor] = None, - negative_prompt_embeds: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + prompt_attention_mask: Optional[torch.Tensor] = None, + negative_prompt_attention_mask: Optional[torch.Tensor] = None, do_classifier_free_guidance: bool = True, max_sequence_length: int = 512, lora_scale: Optional[float] = None, @@ -268,7 +283,7 @@ def encode_prompt( torch device num_images_per_prompt (`int`): number of images that should be generated per prompt - prompt_embeds (`torch.FloatTensor`, *optional*): + prompt_embeds (`torch.Tensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. lora_scale (`float`, *optional*): @@ -293,7 +308,7 @@ def encode_prompt( batch_size = prompt_embeds.shape[0] if prompt_embeds is None: - prompt_embeds = self._get_t5_prompt_embeds( + prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds( prompt=prompt, num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length, @@ -323,12 +338,13 @@ def encode_prompt( " the batch size of `prompt`." ) - negative_prompt_embeds = self._get_t5_prompt_embeds( + negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds( prompt=negative_prompt, num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length, device=device, ) + negative_text_ids = torch.zeros(negative_prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) if self.text_encoder is not None: @@ -336,7 +352,14 @@ def encode_prompt( # Retrieve the original scale by scaling back the LoRA layers unscale_lora_layers(self.text_encoder, lora_scale) - return prompt_embeds, text_ids, negative_prompt_embeds, negative_text_ids + return ( + prompt_embeds, + text_ids, + prompt_attention_mask, + negative_prompt_embeds, + negative_text_ids, + negative_prompt_attention_mask, + ) # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image def encode_image(self, image, device, num_images_per_prompt): @@ -394,7 +417,9 @@ def check_inputs( width, negative_prompt=None, prompt_embeds=None, + prompt_attention_mask=None, negative_prompt_embeds=None, + negative_prompt_attention_mask=None, callback_on_step_end_tensor_inputs=None, max_sequence_length=None, ): @@ -428,6 +453,14 @@ def check_inputs( f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Cannot provide `prompt_embeds` without also providing `prompt_attention_mask") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError( + "Cannot provide `negative_prompt_embeds` without also providing `negative_prompt_attention_mask" + ) + 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}") @@ -534,6 +567,25 @@ def prepare_latents( return latents, latent_image_ids + def _prepare_attention_mask( + self, + batch_size, + sequence_length, + dtype, + attention_mask=None, + ): + if attention_mask is None: + return attention_mask + + # Extend the prompt attention mask to account for image tokens in the final sequence + attention_mask = torch.cat( + [attention_mask, torch.ones(batch_size, sequence_length, device=attention_mask.device)], + dim=1, + ) + attention_mask = attention_mask.to(dtype) + + return attention_mask + @property def guidance_scale(self): return self._guidance_scale @@ -566,18 +618,20 @@ def __call__( negative_prompt: Union[str, List[str]] = None, height: Optional[int] = None, width: Optional[int] = None, - num_inference_steps: int = 28, + num_inference_steps: int = 35, sigmas: Optional[List[float]] = None, - guidance_scale: float = 3.5, + guidance_scale: float = 5.0, 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, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, ip_adapter_image: Optional[PipelineImageInput] = None, ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, negative_ip_adapter_image: Optional[PipelineImageInput] = None, negative_ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, - negative_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + prompt_attention_mask: Optional[torch.Tensor] = None, + negative_prompt_attention_mask: Optional[torch.Tensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, joint_attention_kwargs: Optional[Dict[str, Any]] = None, @@ -618,11 +672,11 @@ def __call__( 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*): + latents (`torch.Tensor`, *optional*): Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will ge generated by sampling using the supplied random `generator`. - prompt_embeds (`torch.FloatTensor`, *optional*): + prompt_embeds (`torch.Tensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. @@ -636,10 +690,18 @@ def __call__( Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. - negative_prompt_embeds (`torch.FloatTensor`, *optional*): + negative_prompt_embeds (`torch.Tensor`, *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. + prompt_attention_mask (torch.Tensor, *optional*): + Attention mask for the prompt embeddings. Used to mask out padding tokens in the prompt sequence. + Chroma requires a single padding token remain unmasked. Please refer to + https://huggingface.co/lodestones/Chroma#tldr-masking-t5-padding-tokens-enhanced-fidelity-and-increased-stability-during-training + negative_prompt_attention_mask (torch.Tensor, *optional*): + Attention mask for the negative prompt embeddings. Used to mask out padding tokens in the negative + prompt sequence. Chroma requires a single padding token remain unmasked. PLease refer to + https://huggingface.co/lodestones/Chroma#tldr-masking-t5-padding-tokens-enhanced-fidelity-and-increased-stability-during-training 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`. @@ -678,7 +740,9 @@ def __call__( width, negative_prompt=negative_prompt, prompt_embeds=prompt_embeds, + prompt_attention_mask=prompt_attention_mask, negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_attention_mask=negative_prompt_attention_mask, callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, max_sequence_length=max_sequence_length, ) @@ -704,13 +768,17 @@ def __call__( ( prompt_embeds, text_ids, + prompt_attention_mask, negative_prompt_embeds, negative_text_ids, + negative_prompt_attention_mask, ) = self.encode_prompt( prompt=prompt, negative_prompt=negative_prompt, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, do_classifier_free_guidance=self.do_classifier_free_guidance, device=device, num_images_per_prompt=num_images_per_prompt, @@ -730,6 +798,7 @@ def __call__( generator, latents, ) + # 5. Prepare timesteps sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas image_seq_len = latents.shape[1] @@ -740,6 +809,20 @@ def __call__( self.scheduler.config.get("base_shift", 0.5), self.scheduler.config.get("max_shift", 1.15), ) + + attention_mask = self._prepare_attention_mask( + batch_size=latents.shape[0], + sequence_length=image_seq_len, + dtype=latents.dtype, + attention_mask=prompt_attention_mask, + ) + negative_attention_mask = self._prepare_attention_mask( + batch_size=latents.shape[0], + sequence_length=image_seq_len, + dtype=latents.dtype, + attention_mask=negative_prompt_attention_mask, + ) + timesteps, num_inference_steps = retrieve_timesteps( self.scheduler, num_inference_steps, @@ -801,6 +884,7 @@ def __call__( encoder_hidden_states=prompt_embeds, txt_ids=text_ids, img_ids=latent_image_ids, + attention_mask=attention_mask, joint_attention_kwargs=self.joint_attention_kwargs, return_dict=False, )[0] @@ -814,6 +898,7 @@ def __call__( encoder_hidden_states=negative_prompt_embeds, txt_ids=negative_text_ids, img_ids=latent_image_ids, + attention_mask=negative_attention_mask, joint_attention_kwargs=self.joint_attention_kwargs, return_dict=False, )[0] diff --git a/src/diffusers/pipelines/chroma/pipeline_chroma_img2img.py b/src/diffusers/pipelines/chroma/pipeline_chroma_img2img.py new file mode 100644 index 000000000000..ce1643147e2b --- /dev/null +++ b/src/diffusers/pipelines/chroma/pipeline_chroma_img2img.py @@ -0,0 +1,1039 @@ +# Copyright 2025 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 CLIPImageProcessor, CLIPVisionModelWithProjection, T5EncoderModel, T5TokenizerFast + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ChromaTransformer2DModel +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 ChromaPipelineOutput + + +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 ChromaTransformer2DModel, ChromaImg2ImgPipeline + >>> from transformers import AutoModel, Autotokenizer + + >>> ckpt_path = "https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v37.safetensors" + >>> transformer = ChromaTransformer2DModel.from_single_file(ckpt_path, torch_dtype=torch.bfloat16) + >>> text_encoder = AutoModel.from_pretrained("black-forest-labs/FLUX.1-schnell", subfolder="text_encoder_2") + >>> tokenizer = AutoTokenizer.from_pretrained("black-forest-labs/FLUX.1-schnell", subfolder="tokenizer_2") + >>> pipe = ChromaImg2ImgPipeline.from_pretrained( + ... "black-forest-labs/FLUX.1-schnell", + ... transformer=transformer, + ... text_encoder=text_encoder, + ... tokenizer=tokenizer, + ... torch_dtype=torch.bfloat16, + ... ) + >>> pipe.enable_model_cpu_offload() + >>> image = load_image( + ... "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + ... ) + >>> prompt = "a scenic fastasy landscape with a river and mountains in the background, vibrant colors, detailed, high resolution" + >>> negative_prompt = "low quality, ugly, unfinished, out of focus, deformed, disfigure, blurry, smudged, restricted palette, flat colors" + >>> image = pipe(prompt, image=image, negative_prompt=negative_prompt).images[0] + >>> image.save("chroma-img2img.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 = 1.15, +): + 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 ChromaImg2ImgPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Chroma pipeline for image-to-image generation. + + Reference: https://huggingface.co/lodestones/Chroma/ + + Args: + transformer ([`ChromaTransformer2DModel`]): + 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 representation + text_encoder ([`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 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: T5EncoderModel, + tokenizer: T5TokenizerFast, + transformer: ChromaTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + + # 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.default_sample_size = 128 + + 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) + + text_inputs = self.tokenizer( + 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 + attention_mask = text_inputs.attention_mask.clone() + + # Chroma requires the attention mask to include one padding token + seq_lengths = attention_mask.sum(dim=1) + mask_indices = torch.arange(attention_mask.size(1)).unsqueeze(0).expand(batch_size, -1) + attention_mask = (mask_indices <= seq_lengths.unsqueeze(1)).long() + + prompt_embeds = self.text_encoder( + text_input_ids.to(device), output_hidden_states=False, attention_mask=attention_mask.to(device) + )[0] + + dtype = self.text_encoder.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + attention_mask = attention_mask.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) + + attention_mask = attention_mask.repeat(1, num_images_per_prompt) + attention_mask = attention_mask.view(batch_size * num_images_per_prompt, seq_len) + + return prompt_embeds, attention_mask + + # 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 + + def encode_prompt( + self, + prompt: Union[str, List[str]], + negative_prompt: Union[str, List[str]] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + prompt_attention_mask: Optional[torch.Tensor] = None, + negative_prompt_attention_mask: Optional[torch.Tensor] = None, + do_classifier_free_guidance: bool = True, + max_sequence_length: int = 512, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `List[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + 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) + + 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_embeds, prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + 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) + negative_text_ids = None + + if do_classifier_free_guidance: + if negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = ( + batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + ) + + 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_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_text_ids = torch.zeros(negative_prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + 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) + + return ( + prompt_embeds, + text_ids, + prompt_attention_mask, + negative_prompt_embeds, + negative_text_ids, + negative_prompt_attention_mask, + ) + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + device = device or self._execution_device + + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def check_inputs( + self, + prompt, + height, + width, + strength, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=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 is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Cannot provide `prompt_embeds` without also providing `prompt_attention_mask") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError( + "Cannot provide `negative_prompt_embeds` without also providing `negative_prompt_attention_mask" + ) + + 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 + def _prepare_latent_image_ids(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 + 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 + 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() + + # 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 + + 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(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) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + 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, latent_image_ids + + def _prepare_attention_mask( + self, + batch_size, + sequence_length, + dtype, + attention_mask=None, + ): + if attention_mask is None: + return attention_mask + + # Extend the prompt attention mask to account for image tokens in the final sequence + attention_mask = torch.cat( + [attention_mask, torch.ones(batch_size, sequence_length, device=attention_mask.device)], + dim=1, + ) + attention_mask = attention_mask.to(dtype) + + return attention_mask + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 35, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 5.0, + strength: float = 0.9, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + negative_ip_adapter_image: Optional[PipelineImageInput] = None, + negative_ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + prompt_attention_mask: Optional[torch.Tensor] = None, + negative_prompt_attention_mask: Optional[torch.tensor] = None, + 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. + 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 + not greater than `1`). + 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. + num_inference_steps (`int`, *optional*, defaults to 35): + 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 5.0): + 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. + strength (`float, *optional*, defaults to 0.9): + Conceptually, indicates how much to transform the reference image. Must be between 0 and 1. image will + be used as a starting point, adding more noise to it the larger the strength. The number of denoising + steps depends on the amount of noise initially added. When strength is 1, added noise will be maximum + and the denoising process will run for the full number of iterations specified in num_inference_steps. + A value of 1, therefore, essentially ignores image. + 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.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_prompt_embeds (`torch.Tensor`, *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. + prompt_attention_mask (torch.Tensor, *optional*): + Attention mask for the prompt embeddings. Used to mask out padding tokens in the prompt sequence. + Chroma requires a single padding token remain unmasked. Please refer to + https://huggingface.co/lodestones/Chroma#tldr-masking-t5-padding-tokens-enhanced-fidelity-and-increased-stability-during-training + negative_prompt_attention_mask (torch.Tensor, *optional*): + Attention mask for the negative prompt embeddings. Used to mask out padding tokens in the negative + prompt sequence. Chroma requires a single padding token remain unmasked. PLease refer to + https://huggingface.co/lodestones/Chroma#tldr-masking-t5-padding-tokens-enhanced-fidelity-and-increased-stability-during-training + 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.ChromaPipelineOutput`] 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.chroma.ChromaPipelineOutput`] or `tuple`: [`~pipelines.chroma.ChromaPipelineOutput`] 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, + height, + width, + strength, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + 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._current_timestep = None + self._interrupt = False + + # 2. Preprocess image + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 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 + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + text_ids, + prompt_attention_mask, + negative_prompt_embeds, + negative_text_ids, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + do_classifier_free_guidance=self.do_classifier_free_guidance, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 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.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + 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) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + 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 // 4 + 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, + ) + + attention_mask = self._prepare_attention_mask( + batch_size=latents.shape[0], + sequence_length=image_seq_len, + dtype=latents.dtype, + attention_mask=prompt_attention_mask, + ) + negative_attention_mask = self._prepare_attention_mask( + batch_size=latents.shape[0], + sequence_length=image_seq_len, + dtype=latents.dtype, + attention_mask=negative_prompt_attention_mask, + ) + + # 6. Prepare image embeddings + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + negative_ip_adapter_image = [negative_ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + ip_adapter_image = [ip_adapter_image] * self.transformer.encoder_hid_proj.num_ip_adapters + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + attention_mask=attention_mask, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + + noise_pred_uncond = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_image_ids, + attention_mask=negative_attention_mask, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred - 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) + + # 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() + + self._current_timestep = None + + 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 ChromaPipelineOutput(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 c95f11fc5848..656a8ac6c6f6 100644 --- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py +++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -272,6 +272,21 @@ def from_pretrained(cls, *args, **kwargs): requires_backends(cls, ["torch", "transformers"]) +class ChromaImg2ImgPipeline(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 ChromaPipeline(metaclass=DummyObject): _backends = ["torch", "transformers"] diff --git a/tests/pipelines/chroma/test_pipeline_chroma_img2img.py b/tests/pipelines/chroma/test_pipeline_chroma_img2img.py new file mode 100644 index 000000000000..02b20527b2f9 --- /dev/null +++ b/tests/pipelines/chroma/test_pipeline_chroma_img2img.py @@ -0,0 +1,170 @@ +import random +import unittest + +import numpy as np +import torch +from transformers import AutoTokenizer, T5EncoderModel + +from diffusers import AutoencoderKL, ChromaImg2ImgPipeline, ChromaTransformer2DModel, FlowMatchEulerDiscreteScheduler +from diffusers.utils.testing_utils import floats_tensor, torch_device + +from ..test_pipelines_common import ( + FluxIPAdapterTesterMixin, + PipelineTesterMixin, + check_qkv_fusion_matches_attn_procs_length, + check_qkv_fusion_processors_exist, +) + + +class ChromaImg2ImgPipelineFastTests( + unittest.TestCase, + PipelineTesterMixin, + FluxIPAdapterTesterMixin, +): + pipeline_class = ChromaImg2ImgPipeline + params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds"]) + batch_params = frozenset(["prompt"]) + + # there is no xformers processor for Flux + test_xformers_attention = False + test_layerwise_casting = True + test_group_offloading = True + + def get_dummy_components(self, num_layers: int = 1, num_single_layers: int = 1): + torch.manual_seed(0) + transformer = ChromaTransformer2DModel( + patch_size=1, + in_channels=4, + num_layers=num_layers, + num_single_layers=num_single_layers, + attention_head_dim=16, + num_attention_heads=2, + joint_attention_dim=32, + axes_dims_rope=[4, 4, 8], + approximator_hidden_dim=32, + approximator_layers=1, + approximator_num_channels=16, + ) + + torch.manual_seed(0) + text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5") + + tokenizer = 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, + "tokenizer": tokenizer, + "transformer": transformer, + "vae": vae, + "image_encoder": None, + "feature_extractor": None, + } + + def get_dummy_inputs(self, device, seed=0): + image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device) + 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, + "height": 8, + "width": 8, + "max_sequence_length": 48, + "strength": 0.8, + "output_type": "np", + } + return inputs + + def test_chroma_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"] = "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_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_chroma_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)