diff --git "a/src/kontext_custom_pipeline.py" "b/src/kontext_custom_pipeline.py" new file mode 100644--- /dev/null +++ "b/src/kontext_custom_pipeline.py" @@ -0,0 +1,3931 @@ +from email.mime import image +import inspect +from turtle import update +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) +from diffusers.image_processor import (VaeImageProcessor, PipelineImageInput) +from diffusers.loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin, FluxIPAdapterMixin +from diffusers.models.autoencoders import AutoencoderKL +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput +from torchvision.transforms.functional import pad +from .transformer_flux import FluxTransformer2DModel +from .detail_encoder import DetailEncoder + +PREFERRED_KONTEXT_RESOLUTIONS = [ + (672, 1568), + (688, 1504), + (720, 1456), + (752, 1392), + (800, 1328), + (832, 1248), + (880, 1184), + (944, 1104), + (1024, 1024), + (1104, 944), + (1184, 880), + (1248, 832), + (1328, 800), + (1392, 752), + (1456, 720), + (1504, 688), + (1568, 672), +] + +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 + +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 + +def prepare_latent_image_ids_2(height, width, device, dtype): + latent_image_ids = torch.zeros(height//2, width//2, 3, device=device, dtype=dtype) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height//2, device=device)[:, None] # y坐标 + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width//2, device=device)[None, :] # x坐标 + return latent_image_ids + +def prepare_latent_subject_ids(height, width, device, dtype): + latent_image_ids = torch.zeros(height // 2, width // 2, 3, device=device, dtype=dtype) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2, device=device)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2, device=device)[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) + +def resize_position_encoding(batch_size, original_height, original_width, target_height, target_width, device, dtype): + latent_image_ids = prepare_latent_image_ids_2(original_height, original_width, device, dtype) + scale_h = original_height / target_height + scale_w = original_width / target_width + 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 + ) + #spatial进行PE插值 + latent_image_ids_resized = torch.zeros(target_height//2, target_width//2, 3, device=device, dtype=dtype) + for i in range(target_height//2): + for j in range(target_width//2): + latent_image_ids_resized[i, j, 1] = i*scale_h + latent_image_ids_resized[i, j, 2] = j*scale_w + cond_latent_image_id_height, cond_latent_image_id_width, cond_latent_image_id_channels = latent_image_ids_resized.shape + cond_latent_image_ids = latent_image_ids_resized.reshape( + cond_latent_image_id_height * cond_latent_image_id_width, cond_latent_image_id_channels + ) + # latent_image_ids_ = torch.concat([latent_image_ids, cond_latent_image_ids], dim=0) + return latent_image_ids, cond_latent_image_ids #, latent_image_ids_ + +# 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, +): + """ + 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 FluxKontextPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux Kontext pipeline for text-to-image generation. + + 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->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: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + detail_encoder: DetailEncoder = None, + ): + 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, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.detail_encoder = detail_encoder + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", 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.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + 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.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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + 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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if 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 negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @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._prepare_latent_image_ids + def _prepare_latent_image_ids_1(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] + height + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + width + + 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 _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], sample_mode="argmax") + 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, sample_mode="argmax") + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_slicing + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_tiling + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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_A: Optional[torch.Tensor], + image_B: Optional[torch.Tensor], + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + ): + if image_B.shape != image_A.shape: + raise ValueError( + f"Image A and Image B must have the same shape, but got {image_A.shape} and {image_B.shape}." + ) + 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) + + image_latents = image_ids = None + if image_A is not None and image_B is not None: + image_A = image_A.to(device=device, dtype=dtype) + image_B = image_B.to(device=device, dtype=dtype) + + if image_A.shape[1] != self.latent_channels: + image_latents_A = self._encode_vae_image(image=image_A, generator=generator) + else: + image_latents_A = image_A + + if image_B.shape[1] != self.latent_channels: + image_latents_B = self._encode_vae_image(image=image_B, generator=generator) + else: + image_latents_B = image_B + + if batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_A.shape[0] + image_latents_A = torch.cat([image_latents_A] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_A` of batch size {image_latents_A.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_A = torch.cat([image_latents_A], dim=0) + + if batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_B.shape[0] + image_latents_B = torch.cat([image_latents_B] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_B` of batch size {image_latents_B.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_B = torch.cat([image_latents_B], dim=0) + + image_latent_height, image_latent_width = image_latents_A.shape[2:] + image_latents_A = self._pack_latents( + image_latents_A, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents_B = self._pack_latents( + image_latents_B, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents = torch.cat([image_latents_A, image_latents_B], dim=1) + + image_ids_A = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + + image_ids_B = self._prepare_latent_image_ids_1( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + image_ids_A[..., 0] = 1 # set first dimension to 0 for image A + image_ids_B[..., 0] = 1 # set first dimension to 1 for image B + image_ids = torch.cat([image_ids_A, image_ids_B], dim=0) + + latent_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents, latent_ids, image_ids + + @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 current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + def __call__( + self, + image_A: Optional[PipelineImageInput] = None, + image_B: Optional[PipelineImageInput] = None, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt: Union[str, List[str]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + true_cfg_scale: float = 1.0, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 3.5, + 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, + 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_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, + max_area: int = 1024**2, + _auto_resize: bool = True, + ): + r""" + Function invoked when calling the pipeline for generation. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_height, original_width = height, width + aspect_ratio = width / height + width = round((max_area * aspect_ratio) ** 0.5) + height = round((max_area / aspect_ratio) ** 0.5) + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + if height != original_height or width != original_width: + logger.warning( + f"Generation `height` and `width` have been adjusted to {height} and {width} to fit the model requirements." + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + 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, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + negative_text_ids, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_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 image + if image_A is not None and image_B is not None and not (isinstance(image_A, torch.Tensor) and image_A.size(1) == self.latent_channels) and not (isinstance(image_B, torch.Tensor) and image_B.size(1) == self.latent_channels): + # reshape image_A as the same size as image_B + img = image_B[0] if isinstance(image_B, list) else image_B + image_height, image_width = self.image_processor.get_default_height_width(img) + aspect_ratio = image_width / image_height + if _auto_resize: + # Kontext is trained on specific resolutions, using one of them is recommended + _, image_width, image_height = min( + (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS + ) + image_width = image_width // multiple_of * multiple_of + image_height = image_height // multiple_of * multiple_of + image_A = self.image_processor.resize(image_A, image_height, image_width) + image_A = self.image_processor.preprocess(image_A, image_height, image_width) + image_B = self.image_processor.resize(image_B, image_height, image_width) + image_B = self.image_processor.preprocess(image_B, image_height, image_width) + else: + raise ValueError("Image input is not supported for costom kontext pipeline.") + + # add step. use the detail_encoder to encoder the image to the prompt_embeds + if self.detail_encoder is not None: + model_dtype = next(self.detail_encoder.parameters()).dtype + model_device = next(self.detail_encoder.parameters()).device + class_token_A_index = 2 + class_token_B_index = 14 + class_tokens_A_mask = [True if class_token_A_index <= i < class_token_A_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_B_mask = [True if class_token_B_index <= i < class_token_B_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_A_mask = torch.tensor(class_tokens_A_mask, dtype=torch.bool).unsqueeze(0) + class_tokens_B_mask = torch.tensor(class_tokens_B_mask, dtype=torch.bool).unsqueeze(0) + A_pixel_values = image_A + B_pixel_values = image_B + A_pixel_values = A_pixel_values.to(device=device, dtype=model_dtype) + A_pixel_values = A_pixel_values.unsqueeze(0) + B_pixel_values = B_pixel_values.to(device=device, dtype=model_dtype) + B_pixel_values = B_pixel_values.unsqueeze(0) + updated_prompt = self.detail_encoder(A_pixel_values, prompt_embeds, class_tokens_A_mask) + updated_prompt = self.detail_encoder(B_pixel_values, updated_prompt, class_tokens_B_mask) + prompt_embeds = updated_prompt + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents, latent_ids, image_ids = self.prepare_latents( + image_A, + image_B, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + if image_ids is not None: + latent_ids = torch.cat([latent_ids, image_ids], dim=0) # dim 0 is sequence dimension + + # 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] + 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, + ) + 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 + + 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 + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + 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 + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + 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_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # 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 FluxPipelineOutput(images=image) + + + + +class FluxKontextPipelineWithPhotoEncoderAddTokens( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux Kontext pipeline for text-to-image generation. + + 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->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: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + detail_encoder: DetailEncoder = None, + ): + 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, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.detail_encoder = detail_encoder + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", 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.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + 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, text_input_ids + + # 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, text_input_ids = 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, text_input_ids + + # 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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + 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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if 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 negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @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 _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], sample_mode="argmax") + 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, sample_mode="argmax") + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_slicing + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_tiling + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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_A: Optional[torch.Tensor], + image_B: Optional[torch.Tensor], + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + ): + if image_B.shape != image_A.shape: + raise ValueError( + f"Image A and Image B must have the same shape, but got {image_A.shape} and {image_B.shape}." + ) + 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) + + image_latents = image_ids = None + if image_A is not None and image_B is not None: + image_A = image_A.to(device=device, dtype=dtype) + image_B = image_B.to(device=device, dtype=dtype) + + if image_A.shape[1] != self.latent_channels: + image_latents_A = self._encode_vae_image(image=image_A, generator=generator) + else: + image_latents_A = image_A + + if image_B.shape[1] != self.latent_channels: + image_latents_B = self._encode_vae_image(image=image_B, generator=generator) + else: + image_latents_B = image_B + + if batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_A.shape[0] + image_latents_A = torch.cat([image_latents_A] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_A` of batch size {image_latents_A.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_A = torch.cat([image_latents_A], dim=0) + + if batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_B.shape[0] + image_latents_B = torch.cat([image_latents_B] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_B` of batch size {image_latents_B.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_B = torch.cat([image_latents_B], dim=0) + + image_latent_height, image_latent_width = image_latents_A.shape[2:] + image_latents_A = self._pack_latents( + image_latents_A, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents_B = self._pack_latents( + image_latents_B, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents = torch.cat([image_latents_A, image_latents_B], dim=1) + + image_ids_A = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + + image_ids_B = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + image_ids_A[..., 0] = 1 # set first dimension to 0 for image A + image_ids_B[..., 0] = 2 # set first dimension to 1 for image B + image_ids = torch.cat([image_ids_A, image_ids_B], dim=0) + + latent_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents, latent_ids, image_ids + + @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 current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + def __call__( + self, + image_A: Optional[PipelineImageInput] = None, + image_B: Optional[PipelineImageInput] = None, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt: Union[str, List[str]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + true_cfg_scale: float = 1.0, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 3.5, + 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, + 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_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, + max_area: int = 1024**2, + _auto_resize: bool = True, + ): + r""" + Function invoked when calling the pipeline for generation. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # original_height, original_width = height, width + # aspect_ratio = width / height + # width = round((max_area * aspect_ratio) ** 0.5) + # height = round((max_area / aspect_ratio) ** 0.5) + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + # if height != original_height or width != original_width: + # logger.warning( + # f"Generation `height` and `width` have been adjusted to {height} and {width} to fit the model requirements." + # ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + trigger_word = ['IMG1', 'IMG2'] + self.tokenizer_2.add_tokens(trigger_word, special_tokens=True) + image_A_token_id = self.tokenizer_2.convert_tokens_to_ids(trigger_word[0]) + image_B_token_id = self.tokenizer_2.convert_tokens_to_ids(trigger_word[1]) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + text_input_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, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + negative_text_ids, + text_input_ids, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_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 image + if image_A is not None and image_B is not None and not (isinstance(image_A, torch.Tensor) and image_A.size(1) == self.latent_channels) and not (isinstance(image_B, torch.Tensor) and image_B.size(1) == self.latent_channels): + # reshape image_A as the same size as image_B + img = image_B[0] if isinstance(image_B, list) else image_B + image_height, image_width = self.image_processor.get_default_height_width(img) + aspect_ratio = image_width / image_height + if _auto_resize: + # Kontext is trained on specific resolutions, using one of them is recommended + _, image_width, image_height = min( + (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS + ) + image_width = image_width // multiple_of * multiple_of + image_height = image_height // multiple_of * multiple_of + image_A = self.image_processor.resize(image_A, image_height, image_width) + image_A = self.image_processor.preprocess(image_A, image_height, image_width) + image_B = self.image_processor.resize(image_B, image_height, image_width) + image_B = self.image_processor.preprocess(image_B, image_height, image_width) + else: + raise ValueError("Image input is not supported for costom kontext pipeline.") + + # add step. use the detail_encoder to encoder the image to the prompt_embeds + if self.detail_encoder is not None: + input_ids = text_input_ids[0] + model_dtype = next(self.detail_encoder.parameters()).dtype + model_device = next(self.detail_encoder.parameters()).device + class_token_A_index = -1 + class_token_B_index = -1 + for index, token_id in enumerate(input_ids.tolist()): + if token_id == image_A_token_id: + class_token_A_index = index + break + for index, token_id in enumerate(input_ids.tolist()): + if token_id == image_B_token_id: + class_token_B_index = index + break + class_tokens_A_mask = [True if class_token_A_index <= i < class_token_A_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_B_mask = [True if class_token_B_index <= i < class_token_B_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_A_mask = torch.tensor(class_tokens_A_mask, dtype=torch.bool).unsqueeze(0) + class_tokens_B_mask = torch.tensor(class_tokens_B_mask, dtype=torch.bool).unsqueeze(0) + A_pixel_values = image_A + B_pixel_values = image_B + A_pixel_values = A_pixel_values.to(device=device, dtype=model_dtype) + A_pixel_values = A_pixel_values.unsqueeze(0) + B_pixel_values = B_pixel_values.to(device=device, dtype=model_dtype) + B_pixel_values = B_pixel_values.unsqueeze(0) + updated_prompt = self.detail_encoder(A_pixel_values, prompt_embeds, class_tokens_A_mask) + updated_prompt = self.detail_encoder(B_pixel_values, updated_prompt, class_tokens_B_mask) + prompt_embeds = updated_prompt + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents, latent_ids, image_ids = self.prepare_latents( + image_A, + image_B, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + if image_ids is not None: + latent_ids = torch.cat([latent_ids, image_ids], dim=0) # dim 0 is sequence dimension + + # 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] + 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, + ) + 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 + + 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 + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + 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 + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + 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_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # 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 FluxPipelineOutput(images=image) + + + +class FluxKontextPipelineWithPhotoEncoderPEeditAddTokens( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux Kontext pipeline for text-to-image generation. + + 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->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: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + detail_encoder: DetailEncoder = None, + ): + 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, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.detail_encoder = detail_encoder + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", 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.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + 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.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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + 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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if 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 negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @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 + def _prepare_latent_image_ids_with_bbox(batch_size, height, width, device, dtype, bboxes=None): + """ + bboxes: Tensor or list with shape (B, 4) where B = batch_size + Each bbox = (x_min, y_min, x_max, y_max) in latent space + """ + assert batch_size == 1 # Currently only supports single batch size + latent_image_ids = torch.zeros(batch_size, height, width, 3, device=device, dtype=dtype) + + y_coords = torch.arange(height, device=device, dtype=dtype).view(1, height, 1) + x_coords = torch.arange(width, device=device, dtype=dtype).view(1, 1, width) + + latent_image_ids[..., 1] = y_coords + latent_image_ids[..., 2] = x_coords + + if bboxes is not None: + if isinstance(bboxes, torch.Tensor): + bboxes = bboxes.tolist() # Convert to list of boxes + if not isinstance(bboxes, list): + bbox_new = [] + bbox_new.append(([bboxes[0], bboxes[1], bboxes[2], bboxes[3]])) + bboxes = bbox_new + + # assert len(bboxes) == batch_size, "bboxes should have shape (B, 4)" + + for b, bbox in enumerate(bboxes): + if len(bbox) != 4: + raise ValueError(f"bbox[{b}] must have 4 elements, got {len(bbox)}") + + x_min, y_min, x_max, y_max = map(int, bbox) + + latent_image_ids[b, y_min:y_max, x_min:x_max, 0] = 1.0 + + bbox_h = y_max - y_min + bbox_w = x_max - x_min + + # y/x in bbox region + y = y_coords[0, y_min:y_max, 0] + x = x_coords[0, 0, x_min:x_max] + + y_scaled = (y - y_min) * (height - 1) / max(bbox_h - 1, 1) + x_scaled = (x - x_min) * (width - 1) / max(bbox_w - 1, 1) + + latent_image_ids[b, y_min:y_max, x_min:x_max, 1] = y_scaled.unsqueeze(1) + latent_image_ids[b, y_min:y_max, x_min:x_max, 2] = x_scaled.unsqueeze(0) + + # Flatten per image to shape (B, H*W, 3) + latent_image_ids = latent_image_ids.view(-1, 3) + return latent_image_ids + + + @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 _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], sample_mode="argmax") + 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, sample_mode="argmax") + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_slicing + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_tiling + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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_A: Optional[torch.Tensor], + image_B: Optional[torch.Tensor], + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + bbox = None, + ): + if image_B.shape != image_A.shape: + raise ValueError( + f"Image A and Image B must have the same shape, but got {image_A.shape} and {image_B.shape}." + ) + 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) + + image_latents = image_ids = None + if image_A is not None and image_B is not None: + image_A = image_A.to(device=device, dtype=dtype) + image_B = image_B.to(device=device, dtype=dtype) + + if image_A.shape[1] != self.latent_channels: + image_latents_A = self._encode_vae_image(image=image_A, generator=generator) + else: + image_latents_A = image_A + + if image_B.shape[1] != self.latent_channels: + image_latents_B = self._encode_vae_image(image=image_B, generator=generator) + else: + image_latents_B = image_B + + if batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_A.shape[0] + image_latents_A = torch.cat([image_latents_A] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_A` of batch size {image_latents_A.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_A = torch.cat([image_latents_A], dim=0) + + if batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_B.shape[0] + image_latents_B = torch.cat([image_latents_B] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_B` of batch size {image_latents_B.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_B = torch.cat([image_latents_B], dim=0) + + image_latent_height, image_latent_width = image_latents_A.shape[2:] + image_latents_A = self._pack_latents( + image_latents_A, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents_B = self._pack_latents( + image_latents_B, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents = torch.cat([image_latents_A, image_latents_B], dim=1) + + image_ids_A = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + + image_ids_B = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + image_ids_A[..., 0] = 1 + image_ids_B[..., 0] = 0 + image_ids = torch.cat([image_ids_A, image_ids_B], dim=0) + + latent_ids = self._prepare_latent_image_ids_with_bbox(batch_size, height // 2, width // 2, device, dtype, bbox) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents, latent_ids, image_ids + + @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 current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + def __call__( + self, + image_A: Optional[PipelineImageInput] = None, + image_B: Optional[PipelineImageInput] = None, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt: Union[str, List[str]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + true_cfg_scale: float = 1.0, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 3.5, + 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, + 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_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, + max_area: int = 1024**2, + _auto_resize: bool = True, + bbox = None, + ): + r""" + Function invoked when calling the pipeline for generation. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_height, original_width = height, width + aspect_ratio = width / height + width = round((max_area * aspect_ratio) ** 0.5) + height = round((max_area / aspect_ratio) ** 0.5) + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + if height != original_height or width != original_width: + logger.warning( + f"Generation `height` and `width` have been adjusted to {height} and {width} to fit the model requirements." + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + 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, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + negative_text_ids, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_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 image + if image_A is not None and image_B is not None and not (isinstance(image_A, torch.Tensor) and image_A.size(1) == self.latent_channels) and not (isinstance(image_B, torch.Tensor) and image_B.size(1) == self.latent_channels): + # reshape image_A as the same size as image_B + img = image_B[0] if isinstance(image_B, list) else image_B + image_height, image_width = self.image_processor.get_default_height_width(img) + aspect_ratio = image_width / image_height + if _auto_resize: + # Kontext is trained on specific resolutions, using one of them is recommended + _, image_width, image_height = min( + (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS + ) + image_width = image_width // multiple_of * multiple_of + image_height = image_height // multiple_of * multiple_of + image_A = self.image_processor.resize(image_A, image_height, image_width) + image_A = self.image_processor.preprocess(image_A, image_height, image_width) + ratio_height = image_height / image_B.size[0] + ratio_width = image_width / image_B.size[1] + bbox = [ + int(bbox[0] * ratio_width), + int(bbox[1] * ratio_height), + int(bbox[2] * ratio_width), + int(bbox[3] * ratio_height) + ] + image_B = self.image_processor.resize(image_B, image_height, image_width) + image_B = self.image_processor.preprocess(image_B, image_height, image_width) + else: + raise ValueError("Image input is not supported for costom kontext pipeline.") + + # add step. use the detail_encoder to encoder the image to the prompt_embeds + if self.detail_encoder is not None: + model_dtype = next(self.detail_encoder.parameters()).dtype + model_device = next(self.detail_encoder.parameters()).device + class_token_A_index = 2 + class_token_B_index = 14 + class_tokens_A_mask = [True if class_token_A_index <= i < class_token_A_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_B_mask = [True if class_token_B_index <= i < class_token_B_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_A_mask = torch.tensor(class_tokens_A_mask, dtype=torch.bool).unsqueeze(0) + class_tokens_B_mask = torch.tensor(class_tokens_B_mask, dtype=torch.bool).unsqueeze(0) + A_pixel_values = image_A + B_pixel_values = image_B + A_pixel_values = A_pixel_values.to(device=device, dtype=model_dtype) + A_pixel_values = A_pixel_values.unsqueeze(0) + B_pixel_values = B_pixel_values.to(device=device, dtype=model_dtype) + B_pixel_values = B_pixel_values.unsqueeze(0) + updated_prompt = self.detail_encoder(A_pixel_values, prompt_embeds, class_tokens_A_mask) + updated_prompt = self.detail_encoder(B_pixel_values, updated_prompt, class_tokens_B_mask) + prompt_embeds = updated_prompt + + # 4. Prepare latent variables + bbox = [ + bbox[0] // 8, + bbox[1] // 8, + bbox[2] // 8, + bbox[3] // 8, + ] + bbox = torch.tensor(bbox).unsqueeze(0).to(device) + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents, latent_ids, image_ids = self.prepare_latents( + image_A, + image_B, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + bbox=bbox, + ) + if image_ids is not None: + latent_ids = torch.cat([latent_ids, image_ids], dim=0) # dim 0 is sequence dimension + + # 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] + 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, + ) + 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 + + 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 + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + 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 + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + 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_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # 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 FluxPipelineOutput(images=image) + + + + +class FluxKontextPipelineWithPhotoEncoderAddTokensVisualization( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux Kontext pipeline for text-to-image generation. + + 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->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: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + detail_encoder: DetailEncoder = None, + ): + 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, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.detail_encoder = detail_encoder + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", 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.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + 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, text_input_ids + + # 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, text_input_ids = 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, text_input_ids + + # 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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + 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 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if 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 negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @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 _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], sample_mode="argmax") + 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, sample_mode="argmax") + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_slicing + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.enable_vae_tiling + 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() + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.disable_vae_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_A: Optional[torch.Tensor], + image_B: Optional[torch.Tensor], + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + ): + if image_B.shape != image_A.shape: + raise ValueError( + f"Image A and Image B must have the same shape, but got {image_A.shape} and {image_B.shape}." + ) + 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) + + image_latents = image_ids = None + if image_A is not None and image_B is not None: + image_A = image_A.to(device=device, dtype=dtype) + image_B = image_B.to(device=device, dtype=dtype) + + if image_A.shape[1] != self.latent_channels: + image_latents_A = self._encode_vae_image(image=image_A, generator=generator) + else: + image_latents_A = image_A + + if image_B.shape[1] != self.latent_channels: + image_latents_B = self._encode_vae_image(image=image_B, generator=generator) + else: + image_latents_B = image_B + + if batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_A.shape[0] + image_latents_A = torch.cat([image_latents_A] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_A.shape[0] and batch_size % image_latents_A.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_A` of batch size {image_latents_A.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_A = torch.cat([image_latents_A], dim=0) + + if batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents_B.shape[0] + image_latents_B = torch.cat([image_latents_B] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents_B.shape[0] and batch_size % image_latents_B.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image_B` of batch size {image_latents_B.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents_B = torch.cat([image_latents_B], dim=0) + + image_latent_height, image_latent_width = image_latents_A.shape[2:] + image_latents_A = self._pack_latents( + image_latents_A, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents_B = self._pack_latents( + image_latents_B, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + image_latents = torch.cat([image_latents_A, image_latents_B], dim=1) + + image_ids_A = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + + image_ids_B = self._prepare_latent_image_ids( + batch_size, image_latent_height // 2, image_latent_width // 2, device, dtype + ) + image_ids_A[..., 0] = 1 # set first dimension to 0 for image A + image_ids_B[..., 0] = 2 # set first dimension to 1 for image B + image_ids = torch.cat([image_ids_A, image_ids_B], dim=0) + + latent_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents, latent_ids, image_ids + + @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 current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + def __call__( + self, + image_A: Optional[PipelineImageInput] = None, + image_B: Optional[PipelineImageInput] = None, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt: Union[str, List[str]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + true_cfg_scale: float = 1.0, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 3.5, + 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, + 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_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, + max_area: int = 1024**2, + _auto_resize: bool = True, + ): + r""" + Function invoked when calling the pipeline for generation. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_height, original_width = height, width + aspect_ratio = width / height + width = round((max_area * aspect_ratio) ** 0.5) + height = round((max_area / aspect_ratio) ** 0.5) + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + if height != original_height or width != original_width: + logger.warning( + f"Generation `height` and `width` have been adjusted to {height} and {width} to fit the model requirements." + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + trigger_word = ['IMG1', 'IMG2'] + self.tokenizer_2.add_tokens(trigger_word, special_tokens=True) + image_A_token_id = self.tokenizer_2.convert_tokens_to_ids(trigger_word[0]) + image_B_token_id = self.tokenizer_2.convert_tokens_to_ids(trigger_word[1]) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + text_input_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, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + negative_text_ids, + text_input_ids, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_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 image + if image_A is not None and image_B is not None and not (isinstance(image_A, torch.Tensor) and image_A.size(1) == self.latent_channels) and not (isinstance(image_B, torch.Tensor) and image_B.size(1) == self.latent_channels): + # reshape image_A as the same size as image_B + img = image_B[0] if isinstance(image_B, list) else image_B + image_height, image_width = self.image_processor.get_default_height_width(img) + aspect_ratio = image_width / image_height + if _auto_resize: + # Kontext is trained on specific resolutions, using one of them is recommended + _, image_width, image_height = min( + (abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS + ) + image_width = image_width // multiple_of * multiple_of + image_height = image_height // multiple_of * multiple_of + image_A = self.image_processor.resize(image_A, image_height, image_width) + image_A = self.image_processor.preprocess(image_A, image_height, image_width) + image_B = self.image_processor.resize(image_B, image_height, image_width) + image_B = self.image_processor.preprocess(image_B, image_height, image_width) + else: + raise ValueError("Image input is not supported for costom kontext pipeline.") + + # add step. use the detail_encoder to encoder the image to the prompt_embeds + if self.detail_encoder is not None: + input_ids = text_input_ids[0] + model_dtype = next(self.detail_encoder.parameters()).dtype + model_device = next(self.detail_encoder.parameters()).device + class_token_A_index = -1 + class_token_B_index = -1 + for index, token_id in enumerate(input_ids.tolist()): + if token_id == image_A_token_id: + class_token_A_index = index + break + for index, token_id in enumerate(input_ids.tolist()): + if token_id == image_B_token_id: + class_token_B_index = index + break + class_tokens_A_mask = [True if class_token_A_index <= i < class_token_A_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_B_mask = [True if class_token_B_index <= i < class_token_B_index+1 else False for i in range(prompt_embeds.shape[1])] + class_tokens_A_mask = torch.tensor(class_tokens_A_mask, dtype=torch.bool).unsqueeze(0) + class_tokens_B_mask = torch.tensor(class_tokens_B_mask, dtype=torch.bool).unsqueeze(0) + A_pixel_values = image_A + B_pixel_values = image_B + A_pixel_values = A_pixel_values.to(device=device, dtype=model_dtype) + A_pixel_values = A_pixel_values.unsqueeze(0) + B_pixel_values = B_pixel_values.to(device=device, dtype=model_dtype) + B_pixel_values = B_pixel_values.unsqueeze(0) + updated_prompt = self.detail_encoder(A_pixel_values, prompt_embeds, class_tokens_A_mask) + updated_prompt = self.detail_encoder(B_pixel_values, updated_prompt, class_tokens_B_mask) + prompt_embeds = updated_prompt + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents, latent_ids, image_ids = self.prepare_latents( + image_A, + image_B, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + if image_ids is not None: + latent_ids = torch.cat([latent_ids, image_ids], dim=0) # dim 0 is sequence dimension + + # 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] + 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, + ) + 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 + + 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 + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + 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 + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + 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_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=negative_text_ids, + img_ids=latent_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # 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 FluxPipelineOutput(images=image) +