import torch from typing import List, Union, Dict, Any, Callable, Optional, Tuple from diffusers.utils.torch_utils import is_compiled_module from diffusers.models import ControlNetModel from diffusers.pipelines.controlnet import MultiControlNetModel from diffusers import StableDiffusionXLControlNetInpaintPipeline from diffusers.image_processor import PipelineImageInput from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput from relighting.pipeline_utils import custom_prepare_latents, custom_prepare_mask_latents, rescale_noise_cfg class CustomStableDiffusionXLControlNetInpaintPipeline(StableDiffusionXLControlNetInpaintPipeline): @torch.no_grad() def __call__( self, prompt: Union[str, List[str]] = None, prompt_2: Optional[Union[str, List[str]]] = None, image: PipelineImageInput = None, mask_image: PipelineImageInput = None, control_image: Union[ PipelineImageInput, List[PipelineImageInput], ] = None, height: Optional[int] = None, width: Optional[int] = None, strength: float = 0.9999, num_inference_steps: int = 50, denoising_start: Optional[float] = None, denoising_end: Optional[float] = None, guidance_scale: float = 5.0, negative_prompt: Optional[Union[str, List[str]]] = None, negative_prompt_2: Optional[Union[str, List[str]]] = None, num_images_per_prompt: Optional[int] = 1, eta: float = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, pooled_prompt_embeds: Optional[torch.FloatTensor] = None, negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, cross_attention_kwargs: Optional[Dict[str, Any]] = None, controlnet_conditioning_scale: Union[float, List[float]] = 1.0, guess_mode: bool = False, control_guidance_start: Union[float, List[float]] = 0.0, control_guidance_end: Union[float, List[float]] = 1.0, guidance_rescale: float = 0.0, original_size: Tuple[int, int] = None, crops_coords_top_left: Tuple[int, int] = (0, 0), target_size: Tuple[int, int] = None, aesthetic_score: float = 6.0, negative_aesthetic_score: float = 2.5, newx: int = 0, newy: int = 0, newr: int = 256, current_seed=0, use_noise_moving=True, ): # OVERWRITE METHODS self.prepare_mask_latents = custom_prepare_mask_latents.__get__(self, CustomStableDiffusionXLControlNetInpaintPipeline) self.prepare_latents = custom_prepare_latents.__get__(self, CustomStableDiffusionXLControlNetInpaintPipeline) controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet # align format for control guidance if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): control_guidance_start = len(control_guidance_end) * [control_guidance_start] elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): control_guidance_end = len(control_guidance_start) * [control_guidance_end] elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [ control_guidance_end ] # # 0.0 Default height and width to unet # height = height or self.unet.config.sample_size * self.vae_scale_factor # width = width or self.unet.config.sample_size * self.vae_scale_factor # 0.1 align format for control guidance if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): control_guidance_start = len(control_guidance_end) * [control_guidance_start] elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): control_guidance_end = len(control_guidance_start) * [control_guidance_end] elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [ control_guidance_end ] # 1. Check inputs self.check_inputs( prompt, prompt_2, control_image, strength, num_inference_steps, callback_steps, negative_prompt, negative_prompt_2, prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, controlnet_conditioning_scale, control_guidance_start, control_guidance_end, ) # 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 # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` # corresponds to doing no classifier free guidance. do_classifier_free_guidance = guidance_scale > 1.0 if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) # 3. Encode input prompt text_encoder_lora_scale = ( cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None ) ( prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, ) = self.encode_prompt( prompt=prompt, prompt_2=prompt_2, device=device, num_images_per_prompt=num_images_per_prompt, do_classifier_free_guidance=do_classifier_free_guidance, 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, lora_scale=text_encoder_lora_scale, ) # 4. set timesteps def denoising_value_valid(dnv): return isinstance(denoising_end, float) and 0 < dnv < 1 self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps, num_inference_steps = self.get_timesteps( num_inference_steps, strength, device, denoising_start=denoising_start if denoising_value_valid else None ) # check that number of inference steps is not < 1 - as this doesn't make sense if num_inference_steps < 1: raise ValueError( f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." ) # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise is_strength_max = strength == 1.0 # 5. Preprocess mask and image - resizes image and mask w.r.t height and width # 5.1 Prepare init image init_image = self.image_processor.preprocess(image, height=height, width=width) init_image = init_image.to(dtype=torch.float32) # 5.2 Prepare control images if isinstance(controlnet, ControlNetModel): control_image = self.prepare_control_image( image=control_image, width=width, height=height, batch_size=batch_size * num_images_per_prompt, num_images_per_prompt=num_images_per_prompt, device=device, dtype=controlnet.dtype, do_classifier_free_guidance=do_classifier_free_guidance, guess_mode=guess_mode, ) elif isinstance(controlnet, MultiControlNetModel): control_images = [] for control_image_ in control_image: control_image_ = self.prepare_control_image( image=control_image_, width=width, height=height, batch_size=batch_size * num_images_per_prompt, num_images_per_prompt=num_images_per_prompt, device=device, dtype=controlnet.dtype, do_classifier_free_guidance=do_classifier_free_guidance, guess_mode=guess_mode, ) control_images.append(control_image_) control_image = control_images else: raise ValueError(f"{controlnet.__class__} is not supported.") # 5.3 Prepare mask mask = self.mask_processor.preprocess(mask_image, height=height, width=width) masked_image = init_image * (mask < 0.5) _, _, height, width = init_image.shape # 6. Prepare latent variables num_channels_latents = self.vae.config.latent_channels num_channels_unet = self.unet.config.in_channels return_image_latents = num_channels_unet == 4 add_noise = True if denoising_start is None else False latents_outputs = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, device, generator, latents, image=init_image, timestep=latent_timestep, is_strength_max=is_strength_max, return_noise=True, return_image_latents=return_image_latents, newx=newx, newy=newy, newr=newr, current_seed=current_seed, use_noise_moving=use_noise_moving, ) if return_image_latents: latents, noise, image_latents = latents_outputs else: latents, noise = latents_outputs # 7. Prepare mask latent variables mask, masked_image_latents = self.prepare_mask_latents( mask, masked_image, batch_size * num_images_per_prompt, height, width, prompt_embeds.dtype, device, generator, do_classifier_free_guidance, ) # 8. Check that sizes of mask, masked image and latents match if num_channels_unet == 9: # default case for runwayml/stable-diffusion-inpainting num_channels_mask = mask.shape[1] num_channels_masked_image = masked_image_latents.shape[1] if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: raise ValueError( f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of" " `pipeline.unet` or your `mask_image` or `image` input." ) elif num_channels_unet != 4: raise ValueError( f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." ) # 8.1 Prepare extra step kwargs. extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) # 8.2 Create tensor stating which controlnets to keep controlnet_keep = [] for i in range(len(timesteps)): keeps = [ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) for s, e in zip(control_guidance_start, control_guidance_end) ] if isinstance(self.controlnet, MultiControlNetModel): controlnet_keep.append(keeps) else: controlnet_keep.append(keeps[0]) # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline height, width = latents.shape[-2:] height = height * self.vae_scale_factor width = width * self.vae_scale_factor original_size = original_size or (height, width) target_size = target_size or (height, width) # 10. Prepare added time ids & embeddings add_text_embeds = pooled_prompt_embeds add_time_ids, add_neg_time_ids = self._get_add_time_ids( original_size, crops_coords_top_left, target_size, aesthetic_score, negative_aesthetic_score, dtype=prompt_embeds.dtype, ) add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) if do_classifier_free_guidance: prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) prompt_embeds = prompt_embeds.to(device) add_text_embeds = add_text_embeds.to(device) add_time_ids = add_time_ids.to(device) # 11. Denoising loop num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) if ( denoising_end is not None and denoising_start is not None and denoising_value_valid(denoising_end) and denoising_value_valid(denoising_start) and denoising_start >= denoising_end ): raise ValueError( f"`denoising_start`: {denoising_start} cannot be larger than or equal to `denoising_end`: " + f" {denoising_end} when using type float." ) elif denoising_end is not None and denoising_value_valid(denoising_end): discrete_timestep_cutoff = int( round( self.scheduler.config.num_train_timesteps - (denoising_end * self.scheduler.config.num_train_timesteps) ) ) num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) timesteps = timesteps[:num_inference_steps] with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents # concat latents, mask, masked_image_latents in the channel dimension latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} # controlnet(s) inference if guess_mode and do_classifier_free_guidance: # Infer ControlNet only for the conditional batch. control_model_input = latents control_model_input = self.scheduler.scale_model_input(control_model_input, t) controlnet_prompt_embeds = prompt_embeds.chunk(2)[1] controlnet_added_cond_kwargs = { "text_embeds": add_text_embeds.chunk(2)[1], "time_ids": add_time_ids.chunk(2)[1], } else: control_model_input = latent_model_input controlnet_prompt_embeds = prompt_embeds controlnet_added_cond_kwargs = added_cond_kwargs if isinstance(controlnet_keep[i], list): cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] else: controlnet_cond_scale = controlnet_conditioning_scale if isinstance(controlnet_cond_scale, list): controlnet_cond_scale = controlnet_cond_scale[0] cond_scale = controlnet_cond_scale * controlnet_keep[i] # # Resize control_image to match the size of the input to the controlnet # if control_image.shape[-2:] != control_model_input.shape[-2:]: # control_image = F.interpolate(control_image, size=control_model_input.shape[-2:], mode="bilinear", align_corners=False) down_block_res_samples, mid_block_res_sample = self.controlnet( control_model_input, t, encoder_hidden_states=controlnet_prompt_embeds, controlnet_cond=control_image, conditioning_scale=cond_scale, guess_mode=guess_mode, added_cond_kwargs=controlnet_added_cond_kwargs, return_dict=False, ) if guess_mode and do_classifier_free_guidance: # Infered ControlNet only for the conditional batch. # To apply the output of ControlNet to both the unconditional and conditional batches, # add 0 to the unconditional batch to keep it unchanged. down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) if num_channels_unet == 9: latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) # predict the noise residual noise_pred = self.unet( latent_model_input, t, encoder_hidden_states=prompt_embeds, cross_attention_kwargs=cross_attention_kwargs, down_block_additional_residuals=down_block_res_samples, mid_block_additional_residual=mid_block_res_sample, added_cond_kwargs=added_cond_kwargs, return_dict=False, )[0] # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) if do_classifier_free_guidance and guidance_rescale > 0.0: print("rescale: ", guidance_rescale) # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] if num_channels_unet == 4: init_latents_proper = image_latents[:1] init_mask = mask[:1] if i < len(timesteps) - 1: noise_timestep = timesteps[i + 1] init_latents_proper = self.scheduler.add_noise( init_latents_proper, noise, torch.tensor([noise_timestep]) ) latents = (1 - init_mask) * init_latents_proper + init_mask * latents # 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 callback is not None and i % callback_steps == 0: callback(i, t, latents) # make sure the VAE is in float32 mode, as it overflows in float16 if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: self.upcast_vae() latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) # If we do sequential model offloading, let's offload unet and controlnet # manually for max memory savings if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: self.unet.to("cpu") self.controlnet.to("cpu") torch.cuda.empty_cache() if not output_type == "latent": image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] else: return StableDiffusionXLPipelineOutput(images=latents) # apply watermark if available if self.watermark is not None: image = self.watermark.apply_watermark(image) image = self.image_processor.postprocess(image, output_type=output_type) # Offload last model to CPU if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: self.final_offload_hook.offload() if not return_dict: return (image,) return StableDiffusionXLPipelineOutput(images=image)