app.py

import os
import torch
import gradio as gr
from tqdm import tqdm
from PIL import Image
from torchvision import transforms as tfms
from transformers import CLIPTextModel, CLIPTokenizer, logging
from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel

torch_device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
if "mps" == torch_device: os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = "1"

# 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")

# 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)

style_token_dict = {'Concept':'<concept-art>', 'Realistic':'<doose-realistic>', 'Line':'<line-art>',
                    'Ricky':'<RickyArt>', 'Plane Scape':'<tony-diterlizzi-planescape>'}

# To the GPU we go!
vae = vae.to(torch_device)
text_encoder = text_encoder.to(torch_device)
unet = unet.to(torch_device)

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_embeddings = pos_emb_layer(position_ids)

concept_art_embed = torch.load('concept-art.bin')
doose_s_realistic_art_style_embed = torch.load('doose-s-realistic-art-style.bin')
line_art_embed = torch.load('line-art.bin')
rickyart_embed = torch.load('rickyart.bin')
tony_diterlizzi_s_planescape_art_embed = torch.load('tony-diterlizzi-s-planescape-art.bin')

tokenizer.add_tokens(['<concept-art>', '<doose-realistic>', '<line-art>', '<RickyArt>', '<tony-diterlizzi-planescape>'])

token_emb_layer_with_art = torch.nn.Embedding(49413, 768)
token_emb_layer_with_art.load_state_dict({'weight': torch.cat((token_emb_layer.state_dict()['weight'],
                                              concept_art_embed['<concept-art>'].unsqueeze(0).to(torch_device),
                                              doose_s_realistic_art_style_embed['<doose-realistic>'].unsqueeze(0).to(torch_device),
                                              line_art_embed['<line-art>'].unsqueeze(0).to(torch_device),
                                              rickyart_embed['<RickyArt>'].unsqueeze(0).to(torch_device),
                                              tony_diterlizzi_s_planescape_art_embed['<tony-diterlizzi-planescape>'].unsqueeze(0).to(torch_device)))})
token_emb_layer_with_art = token_emb_layer_with_art.to(torch_device)

def set_timesteps(scheduler, num_inference_steps):
    scheduler.set_timesteps(num_inference_steps)
    scheduler.timesteps = scheduler.timesteps.to(torch.float32)

def pil_to_latent(input_im):
    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):
    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 build_causal_attention_mask(bsz, seq_len, dtype):
    mask = torch.empty(bsz, seq_len, seq_len, dtype=dtype)
    mask.fill_(torch.tensor(torch.finfo(dtype).min))  # fill with large negative number (acts like -inf)
    mask = mask.triu_(1)  # zero out the lower diagonal to enforce causality
    return mask.unsqueeze(1)  # add a batch dimension

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 = 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

def generate_with_embs(num_inference_steps, guidance_scale, seed, text_input, text_embeddings):
    height = 512                        # default height of Stable Diffusion
    width = 512                         # default width of Stable Diffusion
    generator = torch.manual_seed(seed)   # 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
    set_timesteps(scheduler, num_inference_steps)

    # Prep latents
    latents = torch.randn(
    (batch_size, unet.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), total=len(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 inference(text, style, inference_step, guidance_scale, seed):
    prompt = text + " the style of " + style_token_dict[style]

    # Tokenize
    text_input = tokenizer(prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
    input_ids = text_input.input_ids.to(torch_device)

    # Get token embeddings
    token_embeddings = token_emb_layer_with_art(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:
    image = generate_with_embs(inference_step, guidance_scale, seed, text_input, modified_output_embeddings)

    return image

title = "Stable Diffusion with Textual Inversion"
description = "A simple Gradio interface to infer Stable Diffusion and generate images with different art style"
examples = [["A sweet potato farm", 'Concept', 30, 0.5, 1],
            ["Sky full of cotton candy", 'Realistic', 30, 1.5, 2],
            ["Coffin full of jello", 'Line', 30, 2.5, 3],
            ["Water skiing on a lake", 'Ricky', 30, 3.5, 4],
            ["Super slippery noodles", 'Plane Scape', 30, 4.5, 5],
            ["Beautiful sunset", 'Concept', 30, 5.5, 6],
            ["A glittering gem", 'Realistic', 30, 6.5, 7],
            ["River rafting", 'Line', 30, 7.5, 8],
            ["A green tea", 'Ricky', 30, 8.5, 9],
            ["Three sphered rocks", 'Plane Scape', 30, 9.5, 10]]

demo = gr.Interface(inference, 
                    inputs = [gr.Textbox(label="Prompt", type="text"),
                              gr.Dropdown(label="Style", choices=['Concept', 'Realistic', 'Line', 
                                                                  'Ricky', 'Plane Scape'], value="Concept"), 
                              gr.Slider(10, 50, 30, step = 10, label="Inference steps"),
                              gr.Slider(1, 10, 7.5, step = 0.1, label="Guidance scale"),
                              gr.Slider(0, 10000, 1, step = 1, label="Seed")],
                    outputs= [gr.Image(width=320, height=320, label="Output SAM")],
                    title=title,
                    description=description,
                    examples=examples)

demo.launch()