Code :
# Source : https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wFCHhGLFooW_pf1?usp=sharing#scrollTo=JpjEKYlXXFd0 # !pip install transformers diffusers==0.2.4 # For now specific version needed as update broke something # Parameters # Input image URL # input_image_url = 'https://lafeber.com/pet-birds/wp-content/uploads/2018/06/Scarlet-Macaw-2.jpg' input_image_url = 'https://img.cutenesscdn.com/630x/clsd/getty/33e691dded7d4ddc87a83613469f897e?type=webp' import torch from transformers import CLIPTextModel, CLIPTokenizer from diffusers import AutoencoderKL, UNet2DConditionModel, LMSDiscreteScheduler from tqdm.auto import tqdm from torch import autocast from PIL import Image from matplotlib import pyplot as plt import numpy from torchvision import transforms as tfms # For video display: from IPython.display import HTML from base64 import b64encode import urllib.request # Set device torch_device = "cuda" if torch.cuda.is_available() else "cpu" # Load the autoencoder model which will be used to decode the latents into image space. vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="vae", use_auth_token=False) # Load the tokenizer and text encoder to tokenize and encode the text. tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14") text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14") # The UNet model for generating the latents. unet = UNet2DConditionModel.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="unet", use_auth_token=False) # The noise scheduler scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000) # To the GPU we go! vae = vae.to(torch_device) text_encoder = text_encoder.to(torch_device) unet = unet.to(torch_device) # Using torchvision.transforms.ToTensor to_tensor_tfm = tfms.ToTensor() def pil_to_latent(input_im): # Single image -> single latent in a batch (so size 1, 4, 64, 64) with torch.no_grad(): latent = vae.encode(to_tensor_tfm(input_im).unsqueeze(0).to(torch_device)*2-1) # Note scaling return 0.18215 * latent.mode() # .mode or .mean or .sample # return 0.18215 * latent.latent_dist.mode() # .mode or .mean or .sample def latents_to_pil(latents): # bath of latents -> list of images latents = (1 / 0.18215) * latents with torch.no_grad(): image = vae.decode(latents) # image = 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 # Download a demo Image # !curl --output macaw.jpg 'https://lafeber.com/pet-birds/wp-content/uploads/2018/06/Scarlet-Macaw-2.jpg' # urllib.request.urlretrieve(input_image_url, "alter_input.jpg") # Load the image with PIL input_image = Image.open('alter_input.jpg').resize((512, 512)) # display(input_image) # Encode to the latent space encoded = pil_to_latent(input_image) encoded.shape # Decode this latent representation back into an image decoded = latents_to_pil(encoded)[0] decoded # Setting the number of sampling steps: scheduler.set_timesteps(15) # See these in terms of the original 1000 steps used for training: print(scheduler.timesteps) # Look at the equivalent noise levels: print(scheduler.sigmas) #@markdown Plotting this noise schedule: # plt.plot(scheduler.sigmas) # plt.title('Noise Schedule') # plt.xlabel('Step') # plt.ylabel('sigma') # plt.show() # View a noised version noise = torch.randn_like(encoded) # Random noise timestep = 150 # i.e. equivalent to that at 150/1000 training steps encoded_and_noised = scheduler.add_noise(encoded, noise, timestep) # display(latents_to_pil(encoded_and_noised)[0]) # Display #@title re-generate starting from a noised version of this image prompt = ["A colorful dancer, nat geo photo"] #@param height = 512 # default height of Stable Diffusion width = 512 # default width of Stable Diffusion num_inference_steps = 40 #@param # Number of denoising steps guidance_scale = 8 # Scale for classifier-free guidance generator = torch.manual_seed(32) # Seed generator to create the inital latent noise batch_size = 1 # Prep text text_input = tokenizer(prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") with torch.no_grad(): text_embeddings = text_encoder(text_input.input_ids.to(torch_device))[0] max_length = text_input.input_ids.shape[-1] uncond_input = tokenizer( [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt" ) with torch.no_grad(): uncond_embeddings = text_encoder(uncond_input.input_ids.to(torch_device))[0] text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) # Prep Scheduler scheduler.set_timesteps(num_inference_steps) print("timesteps :") print(scheduler.timesteps) # Start step start_step = 4 #@param Explore ;) start_sigma = scheduler.sigmas[start_step] start_timestep = int(scheduler.timesteps[start_step]) print(f"start_timestep = {start_timestep}") # Prep latents noise = torch.randn_like(encoded) latents = scheduler.add_noise(encoded, noise, start_timestep) # latents = scheduler.add_noise(encoded, noise, timestep) # display(latents_to_pil(latents)[0]) latents = latents.to(torch_device) latents = latents * start_sigma # Loop # with autocast("cuda"): for i, t in tqdm(enumerate(scheduler.timesteps)): if i > start_step: # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes. latent_model_input = torch.cat([latents] * 2) sigma = scheduler.sigmas[i] latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5) # predict the noise residual with torch.no_grad(): noise_pred = 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 + guidance_scale * (noise_pred_text - noise_pred_uncond) # compute the previous noisy sample x_t -> x_t-1 latents = scheduler.step(noise_pred, i, latents)["prev_sample"] latents_to_pil(latents)[0].save("alter.png")
Input image :
Result :