app.py

#!pip install -q --upgrade transformers diffusers ftfy
!pip install -q --upgrade transformers==4.25.1 diffusers ftfy
!pip install accelerate -q

from base64 import b64encode

import numpy
import torch
from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel
from huggingface_hub import notebook_login

# For video display:
from IPython.display import HTML
from matplotlib import pyplot as plt
from pathlib import Path
from PIL import Image
from torch import autocast
from torchvision import transforms as tfms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer, logging

torch.manual_seed(1)
#if not (Path.home()/'.huggingface'/'token').exists(): notebook_login()

# Supress some unnecessary warnings when loading the CLIPTextModel
logging.set_verbosity_error()

# Set device
torch_device = "cuda" if torch.cuda.is_available() else "cpu"

import os
MY_TOKEN=os.environ.get('HF_TOKEN_SD')


# 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=MY_TOKEN)

# 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")

# 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)

"""Functions"""

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(tfms.ToTensor()(input_im).unsqueeze(0).to(torch_device)*2-1) # Note scaling
    return 0.18215 * latent.latent_dist.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).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 get_output_embeds(input_embeddings):
    # CLIP's text model uses causal mask, so we prepare it here:
    bsz, seq_len = input_embeddings.shape[:2]
    causal_attention_mask = 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 = 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(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 = text_encoder.text_model.final_layer_norm(output)

    # And now they're ready!
    return output

#Generating an image with these modified embeddings

def generate_with_embs(text_embeddings, text_input):
    height = 512                        # default height of Stable Diffusion
    width = 512                         # default width of Stable Diffusion
    num_inference_steps = 10            # Number of denoising steps
    guidance_scale = 7.5                # Scale for classifier-free guidance
    generator = torch.manual_seed(64)   # Seed generator to create the inital latent noise
    batch_size = 1

    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)

    # Prep latents
    latents = torch.randn(
    (batch_size, unet.config.in_channels, height // 8, width // 8),
    generator=generator,
    )
    latents = latents.to(torch_device)
    latents = latents * scheduler.init_noise_sigma

    # Loop
    for i, t in tqdm(enumerate(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 = scheduler.sigmas[i]
        latent_model_input = scheduler.scale_model_input(latent_model_input, t)

        # 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, t, latents).prev_sample

    return latents_to_pil(latents)[0]

def ref_loss(images,ref_image):
    # Reference image
    error = torch.abs(images - ref_image).mean()
    return error

def inference(prompt, style_index):

    styles = ['<midjourney-style>', '<hitokomoru-style>','<birb-style>','<summie-style>','<illustration-style>','<m-geo>','<buhu>']
    embed = ['learned_embeds_m.bin','learned_embeds_h.bin', 'learned_embeds.bin',   'learned_embeds_s.bin','learned_embeds_i.bin','learned_embeds_mg.bin','learned_embeds_buhu']


    # Tokenize
    text_input = tokenizer(prompt+" .", padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
    # Access the embedding layer
    token_emb_layer = text_encoder.text_model.embeddings.token_embedding
    token_embeddings = token_emb_layer(text_input.input_ids.to(torch_device))
    pos_emb_layer = text_encoder.text_model.embeddings.position_embedding

    position_ids = text_encoder.text_model.embeddings.position_ids[:, :77]
    position_embeddings = pos_emb_layer(position_ids)

    ## Without any Textual Inversion
    input_ids = text_input.input_ids.to(torch_device)

    # Get token embeddings
    token_embeddings = token_emb_layer(input_ids)

    # Combine with pos embs
    input_embeddings = token_embeddings + position_embeddings

    #  Feed through to get final output embs
    modified_output_embeddings = get_output_embeds(input_embeddings)

    # And generate an image with this:
    image1 = generate_with_embs(modified_output_embeddings,text_input)

    replace_id=269  #replaced dot with Textual Inversion

    ## midjourney-style
    style = styles[style_index]
    emb = embed[style_index]

    x_embed = torch.load(emb)

    # The new embedding - our special birb word
    replacement_token_embedding = x_embed[style].to(torch_device)

    # Insert this into the token embeddings
    token_embeddings[0, torch.where(input_ids[0]==replace_id)] = replacement_token_embedding.to(torch_device)

    # Combine with pos embs
    input_embeddings = token_embeddings + position_embeddings

    #  Feed through to get final output embs
    modified_output_embeddings = get_output_embeds(input_embeddings)

    # And generate an image with this:
    image2 = generate_with_embs(modified_output_embeddings,text_input)

    prompt1 = 'rainbow'

    # Tokenize
    text_input1 = tokenizer(prompt1, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")

    # Access the embedding layer
    token_emb_layer = text_encoder.text_model.embeddings.token_embedding

    pos_emb_layer = text_encoder.text_model.embeddings.position_embedding
    position_ids = text_encoder.text_model.embeddings.position_ids[:, :77]
    position_embeddings1 = pos_emb_layer(position_ids)

    input_ids1 = text_input1.input_ids.to(torch_device)

    # Get token embeddings
    token_embeddings1 = token_emb_layer(input_ids1)

    # Combine with pos embs
    input_embeddings1 = token_embeddings1 + position_embeddings1

    #  Feed through to get final output embs
    modified_output_embeddings1 = get_output_embeds(input_embeddings1)

    # And generate an image with this:
    ref_image = generate_with_embs(modified_output_embeddings1, text_input1)

    ref_latent = pil_to_latent(ref_image)

    height = 512                        # default height of Stable Diffusion
    width = 512                         # default width of Stable Diffusion
    num_inference_steps = 10  #           # Number of denoising steps
    guidance_scale = 8 #               # Scale for classifier-free guidance
    generator = torch.manual_seed(64)   # Seed generator to create the inital latent noise
    batch_size = 1
    blue_loss_scale = 200 #

    # 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]

    # And the uncond. input as before:
    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)

    # Prep latents
    latents = torch.randn(
      (batch_size, unet.config.in_channels, height // 8, width // 8),
      generator=generator,
    )
    latents = latents.to(torch_device)
    latents = latents * scheduler.init_noise_sigma

    # Loop
    for i, t in tqdm(enumerate(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 = scheduler.sigmas[i]
        latent_model_input = scheduler.scale_model_input(latent_model_input, t)

        # predict the noise residual
        with torch.no_grad():
            noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]

        # perform CFG
        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

        #### ADDITIONAL GUIDANCE ###
        if i%5 == 0:
            # Requires grad on the latents
            latents = latents.detach().requires_grad_()

            # Get the predicted x0:
            # latents_x0 = latents - sigma * noise_pred
            latents_x0 = scheduler.step(noise_pred, t, latents).pred_original_sample

            # Decode to image space
            denoised_images = vae.decode((1 / 0.18215) * latents_x0).sample / 2 + 0.5 # range (0, 1)

            #ref image
            with torch.no_grad():
              ref_images = vae.decode((1 / 0.18215) * ref_latent).sample / 2 + 0.5 # range (0, 1)

            # Calculate loss
            loss = ref_loss(denoised_images,ref_images) * blue_loss_scale

            # Occasionally print it out
            # if i%10==0:
            #     print(i, 'loss:', loss.item())

            # Get gradient
            cond_grad = torch.autograd.grad(loss, latents)[0]

            # Modify the latents based on this gradient
            latents = latents.detach() - cond_grad * sigma**2
            scheduler._step_index = scheduler._step_index - 1


        # Now step with scheduler
        latents = scheduler.step(noise_pred, t, latents).prev_sample
        #latents = scheduler.step(noise_pred, t, latents).pred_original_sample


    image3 = latents_to_pil(latents)[0]

    return (image1, 'Original Image'), (image2, 'Styled Image'), (image3, 'After Textual Inversion')

# Gradio App with num_inference_steps=10

title="Textual Inversion in Stable Diffusion"
description="<p style='text-align: center;'>Textual Inversion in Stable Diffusion.</b></p>"
gallery = gr.Gallery(label="Generated images", show_label=True, elem_id="gallery", columns=3).style(grid=[2], height="auto")

gr.Interface(fn=inference, inputs=["text",

    gr.Radio([('<midjourney-style>',0), ('<hitokomoru-style>',1),('<birb-style>',2),
              ('<summie-style>',3),('<illustration-style>',4),('<m-geo>',5),('<buhu>',6)] , value = 0, label = 'Please select your choice of Style')],
    outputs = gallery, title = title,
examples = [['a girl playing in snow',0],
                ['an oil painting of a goddess',6],
                ['a rabbit on the moon', 5 ]], ).launch(debug=True)