import torch from torchvision.transforms import ToTensor from utils import get_style_embeddings, get_EOS_pos_in_prompt, invert_loss from base64 import b64encode import numpy as np from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel from torch import autocast from torchvision import transforms as tfms from tqdm.auto import tqdm from transformers import CLIPTextModel, CLIPTokenizer, logging from PIL import Image import os import torchvision.transforms as T class StableDiffusion: def __init__(self, torch_device, num_inference_steps=30, height=512, width=512, guidance_scale=7.5): # Load the autoencoder vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder='vae') # Load tokenizer and text encoder to tokenize and encode the text self.tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14") text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14") # Unet model for generating latents unet = UNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder='unet') # Noise scheduler self.scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000) # Move everything to GPU self.torch_device = torch_device self.vae = vae.to(self.torch_device) self.text_encoder = text_encoder.to(self.torch_device) self.unet = unet.to(self.torch_device) # additional properties self.num_inference_steps = num_inference_steps self.height = height # default height of Stable Diffusion self.width = width # default width of Stable Diffusion self.guidance_scale = guidance_scale # Scale for classifier-free guidance # Prep Scheduler def set_timesteps(self): self.scheduler.set_timesteps(self.num_inference_steps) self.scheduler.timesteps = self.scheduler.timesteps.to(torch.float32) # minor fix to ensure MPS compatibility, fixed in diffusers PR 3925 def additional_guidance(self, latents, noise_pred, t, sigma, custom_loss_fn, custom_loss_scale): #### ADDITIONAL GUIDANCE ### # Requires grad on the latents latents = latents.detach().requires_grad_() # Get the predicted x0: latents_x0 = latents - sigma * noise_pred #print(f"latents: {latents.shape}, noise_pred:{noise_pred.shape}") #latents_x0 = scheduler.step(noise_pred, t, latents).pred_original_sample # Decode to image space denoised_images = self.vae.decode((1 / 0.18215) * latents_x0).sample / 2 + 0.5 # range (0, 1) # Calculate loss loss = custom_loss_fn(denoised_images) * custom_loss_scale # Get gradient cond_grad = torch.autograd.grad(loss, latents, allow_unused=False)[0] # Modify the latents based on this gradient latents = latents.detach() - cond_grad * sigma**2 return latents, loss def generate_with_embs(self, text_embeddings, max_length, random_seed, custom_loss_fn, custom_loss_scale): generator = torch.manual_seed(random_seed) # Seed generator to create the inital latent noise batch_size = 1 uncond_input = self.tokenizer( [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt" ) with torch.no_grad(): uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.torch_device))[0] text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) # Prep Scheduler self.set_timesteps() # Prep latents latents = torch.randn( (batch_size, self.unet.in_channels, self.height // 8, self.width // 8), generator=generator,) latents = latents.to(self.torch_device) latents = latents * self.scheduler.init_noise_sigma # Loop for i, t in tqdm(enumerate(self.scheduler.timesteps), total=len(self.scheduler.timesteps)): # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes. latent_model_input = torch.cat([latents] * 2) sigma = self.scheduler.sigmas[i] latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) # predict the noise residual with torch.no_grad(): noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"] # perform guidance noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) if custom_loss_fn is not None: if i%10 == 0: latents, custom_loss = self.additional_guidance(latents, noise_pred, t, sigma, custom_loss_fn, custom_loss_scale) print(i, 'loss:', custom_loss.item()) # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents).prev_sample return self.latents_to_pil(latents)[0] def get_output_embeds(self, input_embeddings): # CLIP's text model uses causal mask, so we prepare it here: bsz, seq_len = input_embeddings.shape[:2] causal_attention_mask = self.text_encoder.text_model._build_causal_attention_mask(bsz, seq_len, dtype=input_embeddings.dtype) # Getting the output embeddings involves calling the model with passing output_hidden_states=True # so that it doesn't just return the pooled final predictions: encoder_outputs = self.text_encoder.text_model.encoder( inputs_embeds=input_embeddings, attention_mask=None, # We aren't using an attention mask so that can be None causal_attention_mask=causal_attention_mask.to(self.torch_device), output_attentions=None, output_hidden_states=True, # We want the output embs not the final output return_dict=None, ) # We're interested in the output hidden state only output = encoder_outputs[0] # There is a final layer norm we need to pass these through output = self.text_encoder.text_model.final_layer_norm(output) # And now they're ready! return output def pil_to_latent(self, input_im): # Single image -> single latent in a batch (so size 1, 4, 64, 64) with torch.no_grad(): latent = self.vae.encode(tfms.ToTensor()(input_im).unsqueeze(0).to(self.torch_device)*2-1) # Note scaling return 0.18215 * latent.latent_dist.sample() def latents_to_pil(self, latents): # bath of latents -> list of images latents = (1 / 0.18215) * latents with torch.no_grad(): image = self.vae.decode(latents).sample image = (image / 2 + 0.5).clamp(0, 1) image = image.detach().cpu().permute(0, 2, 3, 1).numpy() images = (image * 255).round().astype("uint8") pil_images = [Image.fromarray(image) for image in images] return pil_images def generate_image_with_custom_style(self, prompt, style_token_embedding=None, random_seed=41, custom_loss_fn = None, custom_loss_scale=None): eos_pos = get_EOS_pos_in_prompt(prompt) # tokenize text_input = self.tokenizer(prompt, padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt") max_length = text_input.input_ids.shape[-1] input_ids = text_input.input_ids.to(self.torch_device) # get token embeddings token_emb_layer = self.text_encoder.text_model.embeddings.token_embedding token_embeddings = token_emb_layer(input_ids) # Append style token towards the end of the sentence embeddings if style_token_embedding is not None: token_embeddings[-1, eos_pos, :] = style_token_embedding # combine with pos embs pos_emb_layer = self.text_encoder.text_model.embeddings.position_embedding position_ids = self.text_encoder.text_model.embeddings.position_ids[:, :77] position_embeddings = pos_emb_layer(position_ids) input_embeddings = token_embeddings + position_embeddings # Feed through to get final output embs modified_output_embeddings = self.get_output_embeds(input_embeddings) # And generate an image with this: generated_image = self.generate_with_embs(modified_output_embeddings, max_length, random_seed, custom_loss_fn, custom_loss_scale) return generated_image