import torch import numpy as np from typing import Any, Callable, Dict, Tuple, List, Optional, Union from diffusers import FluxTransformer2DModel from diffusers.configuration_utils import register_to_config from diffusers.utils import logging, USE_PEFT_BACKEND, scale_lora_layers, unscale_lora_layers from diffusers.models.modeling_outputs import Transformer2DModelOutput from diffusers.pipelines.flux.pipeline_flux import FluxPipeline, calculate_shift, retrieve_timesteps from diffusers.image_processor import PipelineImageInput from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput logger = logging.get_logger(__name__) # pylint: disable=invalid-name def drop_token(x, drop_prob: float = 0., training: bool = False, scale_by_keep: bool = True): if drop_prob == 0. or not training: return x keep_prob = 1 - drop_prob shape = (x.shape[0], x.shape[1], 1) random_tensor = x.new_empty(shape).bernoulli_(keep_prob) if keep_prob > 0.0 and scale_by_keep: random_tensor.div_(keep_prob) return x * random_tensor class FluxTransformer2DModelWithSigLIP(FluxTransformer2DModel): @register_to_config def __init__( self, patch_size: int = 1, in_channels: int = 64, out_channels: Optional[int] = None, num_layers: int = 19, num_single_layers: int = 38, attention_head_dim: int = 128, num_attention_heads: int = 24, joint_attention_dim: int = 4096, pooled_projection_dim: int = 768, guidance_embeds: bool = False, axes_dims_rope: Tuple[int] = (16, 56, 56), siglip_channels: Optional[int] = None, drop_token_prob: float = 0., ): super().__init__( patch_size=patch_size, in_channels=in_channels, out_channels=out_channels, num_layers=num_layers, num_single_layers=num_single_layers, attention_head_dim=attention_head_dim, num_attention_heads=num_attention_heads, joint_attention_dim=joint_attention_dim, pooled_projection_dim=pooled_projection_dim, guidance_embeds=guidance_embeds, axes_dims_rope=axes_dims_rope, ) self.drop_token_prob = drop_token_prob if siglip_channels is not None: self.init_siglip_embed(siglip_channels) def init_siglip_embed(self, siglip_channels): self.siglip_embed = torch.nn.Linear(siglip_channels, self.inner_dim, bias=False) torch.nn.init.zeros_(self.siglip_embed.weight) def forward( self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor = None, pooled_projections: torch.Tensor = None, timestep: torch.LongTensor = None, img_ids: torch.Tensor = None, txt_ids: torch.Tensor = None, guidance: torch.Tensor = None, siglip_tensor: Optional[torch.Tensor] = None, joint_attention_kwargs: Optional[Dict[str, Any]] = None, controlnet_block_samples=None, controlnet_single_block_samples=None, return_dict: bool = True, controlnet_blocks_repeat: bool = False, ) -> Union[torch.Tensor, Transformer2DModelOutput]: """ The [`FluxTransformer2DModel`] forward method. Args: hidden_states (`torch.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`): Input `hidden_states`. encoder_hidden_states (`torch.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`): Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. pooled_projections (`torch.Tensor` of shape `(batch_size, projection_dim)`): Embeddings projected from the embeddings of input conditions. timestep ( `torch.LongTensor`): Used to indicate denoising step. block_controlnet_hidden_states: (`list` of `torch.Tensor`): A list of tensors that if specified are added to the residuals of transformer blocks. joint_attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain tuple. Returns: If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a `tuple` where the first element is the sample tensor. """ if joint_attention_kwargs is not None: joint_attention_kwargs = joint_attention_kwargs.copy() lora_scale = joint_attention_kwargs.pop("scale", 1.0) else: lora_scale = 1.0 if USE_PEFT_BACKEND: # weight the lora layers by setting `lora_scale` for each PEFT layer scale_lora_layers(self, lora_scale) else: if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None: logger.warning( "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective." ) hidden_states = self.x_embedder(hidden_states) timestep = timestep.to(hidden_states.dtype) * 1000 if guidance is not None: guidance = guidance.to(hidden_states.dtype) * 1000 else: guidance = None temb = ( self.time_text_embed(timestep, pooled_projections) if guidance is None else self.time_text_embed(timestep, guidance, pooled_projections) ) encoder_hidden_states = self.context_embedder(encoder_hidden_states) if txt_ids.ndim == 3: logger.warning( "Passing `txt_ids` 3d torch.Tensor is deprecated." "Please remove the batch dimension and pass it as a 2d torch Tensor" ) txt_ids = txt_ids[0] if img_ids.ndim == 3: logger.warning( "Passing `img_ids` 3d torch.Tensor is deprecated." "Please remove the batch dimension and pass it as a 2d torch Tensor" ) img_ids = img_ids[0] ids = torch.cat((txt_ids, img_ids), dim=0) image_rotary_emb = self.pos_embed(ids) if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs: ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds") ip_hidden_states = self.encoder_hid_proj(ip_adapter_image_embeds) joint_attention_kwargs.update({"ip_hidden_states": ip_hidden_states}) for index_block, block in enumerate(self.transformer_blocks): if torch.is_grad_enabled() and self.gradient_checkpointing: encoder_hidden_states, hidden_states = self._gradient_checkpointing_func( block, hidden_states, encoder_hidden_states, temb, image_rotary_emb, ) else: encoder_hidden_states, hidden_states = block( hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, joint_attention_kwargs=joint_attention_kwargs, ) # controlnet residual if controlnet_block_samples is not None: interval_control = len(self.transformer_blocks) / len(controlnet_block_samples) interval_control = int(np.ceil(interval_control)) # For Xlabs ControlNet. if controlnet_blocks_repeat: hidden_states = ( hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)] ) else: hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control] if siglip_tensor is not None: siglip_tensor = drop_token(siglip_tensor, self.drop_token_prob, training=self.training) hidden_states = hidden_states + self.siglip_embed(siglip_tensor) hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) for index_block, block in enumerate(self.single_transformer_blocks): if torch.is_grad_enabled() and self.gradient_checkpointing: hidden_states = self._gradient_checkpointing_func( block, hidden_states, temb, image_rotary_emb, ) else: hidden_states = block( hidden_states=hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, joint_attention_kwargs=joint_attention_kwargs, ) # controlnet residual if controlnet_single_block_samples is not None: interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples) interval_control = int(np.ceil(interval_control)) hidden_states[:, encoder_hidden_states.shape[1]:, ...] = ( hidden_states[:, encoder_hidden_states.shape[1]:, ...] + controlnet_single_block_samples[index_block // interval_control] ) hidden_states = hidden_states[:, encoder_hidden_states.shape[1]:, ...] hidden_states = self.norm_out(hidden_states, temb) output = self.proj_out(hidden_states) if USE_PEFT_BACKEND: # remove `lora_scale` from each PEFT layer unscale_lora_layers(self, lora_scale) if not return_dict: return (output,) return Transformer2DModelOutput(sample=output) def teacache_forward( self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor = None, pooled_projections: torch.Tensor = None, timestep: torch.LongTensor = None, img_ids: torch.Tensor = None, txt_ids: torch.Tensor = None, guidance: torch.Tensor = None, joint_attention_kwargs: Optional[Dict[str, Any]] = None, controlnet_block_samples=None, controlnet_single_block_samples=None, return_dict: bool = True, controlnet_blocks_repeat: bool = False, siglip_tensor: Optional[torch.Tensor] = None, ) -> Union[torch.FloatTensor, Transformer2DModelOutput]: """ The [`FluxTransformer2DModel`] forward method. Args: hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input `hidden_states`. encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`): Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected from the embeddings of input conditions. timestep ( `torch.LongTensor`): Used to indicate denoising step. block_controlnet_hidden_states: (`list` of `torch.Tensor`): A list of tensors that if specified are added to the residuals of transformer blocks. joint_attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain tuple. Returns: If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a `tuple` where the first element is the sample tensor. """ if joint_attention_kwargs is not None: joint_attention_kwargs = joint_attention_kwargs.copy() lora_scale = joint_attention_kwargs.pop("scale", 1.0) else: lora_scale = 1.0 if USE_PEFT_BACKEND: # weight the lora layers by setting `lora_scale` for each PEFT layer scale_lora_layers(self, lora_scale) else: if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None: logger.warning( "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective." ) batch_size, seq_len, channels = hidden_states.shape device, dtype = hidden_states.device, hidden_states.dtype hidden_states = self.x_embedder(hidden_states) timestep = timestep.to(hidden_states.dtype) * 1000 if guidance is not None: guidance = guidance.to(hidden_states.dtype) * 1000 else: guidance = None temb = ( self.time_text_embed(timestep, pooled_projections) if guidance is None else self.time_text_embed(timestep, guidance, pooled_projections) ) encoder_hidden_states = self.context_embedder(encoder_hidden_states) if txt_ids.ndim == 3: logger.warning( "Passing `txt_ids` 3d torch.Tensor is deprecated." "Please remove the batch dimension and pass it as a 2d torch Tensor" ) txt_ids = txt_ids[0] if img_ids.ndim == 3: logger.warning( "Passing `img_ids` 3d torch.Tensor is deprecated." "Please remove the batch dimension and pass it as a 2d torch Tensor" ) img_ids = img_ids[0] ids = torch.cat((txt_ids, img_ids), dim=0) image_rotary_emb = self.pos_embed(ids) if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs: ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds") ip_hidden_states = self.encoder_hid_proj(ip_adapter_image_embeds) joint_attention_kwargs.update({"ip_hidden_states": ip_hidden_states}) if self.enable_teacache: inp = hidden_states.clone() temb_ = temb.clone() modulated_inp, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.transformer_blocks[0].norm1(inp, emb=temb_) if self.cnt == 0 or self.cnt == self.num_steps - 1: should_calc = True self.accumulated_rel_l1_distance = 0 else: coefficients = [4.98651651e+02, -2.83781631e+02, 5.58554382e+01, -3.82021401e+00, 2.64230861e-01] rescale_func = np.poly1d(coefficients) # rescale_func = Polynomial(coefficients.reverse()) self.accumulated_rel_l1_distance += rescale_func(((modulated_inp - self.previous_modulated_input).abs().mean() / self.previous_modulated_input.abs().mean()).cpu().item()) if self.accumulated_rel_l1_distance < self.rel_l1_thresh: should_calc = False else: should_calc = True self.accumulated_rel_l1_distance = 0 self.previous_modulated_input = modulated_inp self.cnt += 1 if self.cnt == self.num_steps: self.cnt = 0 if self.enable_teacache: if not should_calc: hidden_states += self.previous_residual else: ori_hidden_states = hidden_states.clone() for index_block, block in enumerate(self.transformer_blocks): if torch.is_grad_enabled() and self.gradient_checkpointing: encoder_hidden_states, hidden_states = self._gradient_checkpointing_func( block, hidden_states, encoder_hidden_states, temb, image_rotary_emb, ) else: encoder_hidden_states, hidden_states = block( hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, joint_attention_kwargs=joint_attention_kwargs, ) # controlnet residual if controlnet_block_samples is not None: interval_control = len(self.transformer_blocks) / len(controlnet_block_samples) interval_control = int(np.ceil(interval_control)) # For Xlabs ControlNet. if controlnet_blocks_repeat: hidden_states = ( hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)] ) else: hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control] if siglip_tensor is not None: siglip_tensor = drop_token(siglip_tensor, self.drop_token_prob, training=self.training) hidden_states = hidden_states + self.siglip_embed(siglip_tensor) hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) for index_block, block in enumerate(self.single_transformer_blocks): if torch.is_grad_enabled() and self.gradient_checkpointing: hidden_states = self._gradient_checkpointing_func( block, hidden_states, temb, image_rotary_emb, ) else: hidden_states = block( hidden_states=hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, joint_attention_kwargs=joint_attention_kwargs, ) # controlnet residual if controlnet_single_block_samples is not None: interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples) interval_control = int(np.ceil(interval_control)) hidden_states[:, encoder_hidden_states.shape[1]:, ...] = ( hidden_states[:, encoder_hidden_states.shape[1]:, ...] + controlnet_single_block_samples[index_block // interval_control] ) hidden_states = hidden_states[:, encoder_hidden_states.shape[1]:, ...] self.previous_residual = hidden_states - ori_hidden_states else: for index_block, block in enumerate(self.transformer_blocks): if torch.is_grad_enabled() and self.gradient_checkpointing: encoder_hidden_states, hidden_states = self._gradient_checkpointing_func( block, hidden_states, encoder_hidden_states, temb, image_rotary_emb, ) else: encoder_hidden_states, hidden_states = block( hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, joint_attention_kwargs=joint_attention_kwargs, ) # controlnet residual if controlnet_block_samples is not None: interval_control = len(self.transformer_blocks) / len(controlnet_block_samples) interval_control = int(np.ceil(interval_control)) # For Xlabs ControlNet. if controlnet_blocks_repeat: hidden_states = ( hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)] ) else: hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control] if siglip_tensor is not None: siglip_tensor = drop_token(siglip_tensor, self.drop_token_prob, training=self.training) hidden_states = hidden_states + self.siglip_embed(siglip_tensor) hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) for index_block, block in enumerate(self.single_transformer_blocks): if torch.is_grad_enabled() and self.gradient_checkpointing: hidden_states = self._gradient_checkpointing_func( block, hidden_states, temb, image_rotary_emb, ) else: hidden_states = block( hidden_states=hidden_states, temb=temb, image_rotary_emb=image_rotary_emb, joint_attention_kwargs=joint_attention_kwargs, ) # controlnet residual if controlnet_single_block_samples is not None: interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples) interval_control = int(np.ceil(interval_control)) hidden_states[:, encoder_hidden_states.shape[1]:, ...] = ( hidden_states[:, encoder_hidden_states.shape[1]:, ...] + controlnet_single_block_samples[index_block // interval_control] ) hidden_states = hidden_states[:, encoder_hidden_states.shape[1]:, ...] hidden_states = self.norm_out(hidden_states, temb) output = self.proj_out(hidden_states) if USE_PEFT_BACKEND: # remove `lora_scale` from each PEFT layer unscale_lora_layers(self, lora_scale) if not return_dict: return (output,) return Transformer2DModelOutput(sample=output) class FluxPipelineWithSigLIP(FluxPipeline): @torch.no_grad() def __call__( self, siglip_tensor: torch.Tensor, 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, true_cfg_scale_2: 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, ): r""" Function invoked when calling the pipeline for generation. Args: prompt (`str` or `List[str]`, *optional*): The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. instead. prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is will be used instead. negative_prompt (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is not greater than `1`). negative_prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. true_cfg_scale (`float`, *optional*, defaults to 1.0): When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): The height in pixels of the generated image. This is set to 1024 by default for the best results. width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): The width in pixels of the generated image. This is set to 1024 by default for the best results. num_inference_steps (`int`, *optional*, defaults to 50): The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. sigmas (`List[float]`, *optional*): Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed will be used. guidance_scale (`float`, *optional*, defaults to 3.5): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. generator (`torch.Generator` or `List[torch.Generator]`, *optional*): One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation deterministic. latents (`torch.FloatTensor`, *optional*): Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will ge generated by sampling using the supplied random `generator`. prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, pooled text embeddings will be generated from `prompt` input argument. ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. negative_ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. negative_ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. negative_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` input argument. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generate image. Choose between [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. joint_attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). callback_on_step_end (`Callable`, *optional*): A function that calls at the end of each denoising steps during the inference. The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by `callback_on_step_end_tensor_inputs`. callback_on_step_end_tensor_inputs (`List`, *optional*): The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the `._callback_tensor_inputs` attribute of your pipeline class. max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. Examples: Returns: [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. """ assert true_cfg_scale == true_cfg_scale_2 height = height or self.default_sample_size * self.vae_scale_factor width = width or self.default_sample_size * self.vae_scale_factor # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, prompt_2, 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, ) assert do_true_cfg ( negative_prompt_embeds, negative_pooled_prompt_embeds, _, ) = 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, ) # 4. Prepare latent variables num_channels_latents = self.transformer.config.in_channels // 4 latents, latent_image_ids = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, device, generator, latents, ) # 5. Prepare timesteps sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas image_seq_len = latents.shape[1] 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] * 2) 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 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 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML timestep = t.expand(latents.shape[0] * 2).to(latents.dtype) batch_noise_pred = self.transformer( hidden_states=torch.cat([latents, latents], dim=0), timestep=timestep / 1000, guidance=guidance, pooled_projections=torch.cat([pooled_prompt_embeds, negative_pooled_prompt_embeds.expand_as(pooled_prompt_embeds)], dim=0), encoder_hidden_states=torch.cat([prompt_embeds, negative_prompt_embeds.expand_as(prompt_embeds)], dim=0), txt_ids=text_ids, img_ids=latent_image_ids, joint_attention_kwargs=self.joint_attention_kwargs, siglip_tensor=torch.cat([siglip_tensor, torch.zeros_like(siglip_tensor)], dim=0), return_dict=False, )[0] noise_pred, neg_noise_pred = batch_noise_pred.chunk(2) 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() 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)