app.py

import gradio as gr

import torch
from torch import autocast

import gc
import io
import math
import sys

from PIL import Image, ImageOps
import requests
from torch import nn
from torch.nn import functional as F
from torchvision import transforms
from torchvision.transforms import functional as TF
from tqdm.notebook import tqdm

import numpy as np

from guided_diffusion.script_util import create_model_and_diffusion, model_and_diffusion_defaults, classifier_defaults, create_classifier

from omegaconf import OmegaConf
from ldm.util import instantiate_from_config

from einops import rearrange
from math import log2, sqrt

import argparse
import pickle

import os

from transformers import CLIPTokenizer, CLIPTextModel

def fetch(url_or_path):
    if str(url_or_path).startswith('http://') or str(url_or_path).startswith('https://'):
        r = requests.get(url_or_path)
        r.raise_for_status()
        fd = io.BytesIO()
        fd.write(r.content)
        fd.seek(0)
        return fd
    return open(url_or_path, 'rb')

device = "cuda"

#model_state_dict = torch.load('diffusion.pt', map_location='cpu')
model_state_dict = torch.load(fetch('https://huggingface.co./Jack000/glid-3-xl-stable/resolve/main/default/diffusion-1.4.pt'), map_location='cpu')

model_params = {
    'attention_resolutions': '32,16,8',
    'class_cond': False,
    'diffusion_steps': 1000,
    'rescale_timesteps': True,
    'timestep_respacing': 'ddim100',
    'image_size': 32,
    'learn_sigma': False,
    'noise_schedule': 'linear',
    'num_channels': 320,
    'num_heads': 8,
    'num_res_blocks': 2,
    'resblock_updown': False,
    'use_fp16': True,
    'use_scale_shift_norm': False,
    'clip_embed_dim': None,
    'image_condition': False,
    'super_res_condition': False,
}

model_config = model_and_diffusion_defaults()
model_config.update(model_params)

# Load models
model, diffusion = create_model_and_diffusion(**model_config)
model.load_state_dict(model_state_dict, strict=True)
model.requires_grad_(False).eval().to(device)

if model_config['use_fp16']:
    model.convert_to_fp16()
else:
    model.convert_to_fp32()

def set_requires_grad(model, value):
    for param in model.parameters():
        param.requires_grad = value

# vae
kl_config = OmegaConf.load('kl.yaml')
kl_sd = torch.load(fetch('https://huggingface.co./Jack000/glid-3-xl-stable/resolve/main/default/kl-1.4.pt'), map_location="cpu")

ldm = instantiate_from_config(kl_config.model)
ldm.load_state_dict(kl_sd, strict=True)

ldm.to(device)
ldm.eval()
ldm.requires_grad_(False)
set_requires_grad(ldm, False)

# clip
clip_version = 'openai/clip-vit-large-patch14'
clip_tokenizer = CLIPTokenizer.from_pretrained(clip_version)
clip_transformer = CLIPTextModel.from_pretrained(clip_version)
clip_transformer.eval().requires_grad_(False).to(device)

# classifier
# load classifier
classifier_config = classifier_defaults()
classifier_config['classifier_width'] = 128
classifier_config['classifier_depth'] = 4
classifier_config['classifier_attention_resolutions'] = '64,32,16,8'

classifier_photo = create_classifier(**classifier_config)
classifier_photo.load_state_dict(
    torch.load(fetch('https://huggingface.co./Jack000/glid-3-xl-stable/resolve/main/classifier_photo/model060000.pt'), map_location="cpu")
)
classifier_photo.to(device)
classifier_photo.convert_to_fp16()
classifier_photo.eval()

classifier_art = create_classifier(**classifier_config)
classifier_art.load_state_dict(
    torch.load(fetch('https://huggingface.co./Jack000/glid-3-xl-stable/resolve/main/classifier_art/model110000.pt'), map_location="cpu")
)
classifier_art.to(device)
classifier_art.convert_to_fp16()
classifier_art.eval()

def infer(prompt, style, scale, classifier_scale, seed):
    torch.manual_seed(seed)

    # clip context
    text = clip_tokenizer([prompt], truncation=True, max_length=77, return_length=True, return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
    text_blank = clip_tokenizer([''], truncation=True, max_length=77, return_length=True, return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
    text_tokens = text["input_ids"].to(device)
    text_blank_tokens = text_blank["input_ids"].to(device)

    text_emb = clip_transformer(input_ids=text_tokens).last_hidden_state
    text_emb_blank = clip_transformer(input_ids=text_blank_tokens).last_hidden_state

    kwargs = {
        "context": torch.cat([text_emb, text_emb_blank], dim=0).half(),
        "clip_embed": None,
        "image_embed": None,
    }

    def model_fn(x_t, ts, **kwargs):
        half = x_t[: len(x_t) // 2]
        combined = torch.cat([half, half], dim=0)
        model_out = model(combined, ts, **kwargs)
        eps, rest = model_out[:, :3], model_out[:, 3:]
        cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
        half_eps = uncond_eps + scale * (cond_eps - uncond_eps)
        eps = torch.cat([half_eps, half_eps], dim=0)
        return torch.cat([eps, rest], dim=1)

    def cond_fn(x, t, context=None, clip_embed=None, image_embed=None):
        with torch.enable_grad():
            x_in = x[:x.shape[0]//2].detach().requires_grad_(True)
            if style == 'photo':
                logits = classifier_photo(x_in, t)
            elif style == 'digital art':
                logits = classifier_art(x_in, t)
            else:
                return 0

            log_probs = F.log_softmax(logits, dim=-1)
            selected = log_probs[range(len(logits)), torch.ones(x_in.shape[0], dtype=torch.long)]
            return torch.autograd.grad(selected.sum(), x_in)[0] * classifier_scale

    samples = diffusion.ddim_sample_loop_progressive(
        model_fn,
        (2, 4, 64, 64),
        clip_denoised=False,
        model_kwargs=kwargs,
        cond_fn=cond_fn,
        device=device,
        progress=True,
        init_image=None,
        skip_timesteps=0,
    )

    for j, sample in enumerate(samples):
        pass

    emb = sample['pred_xstart'][0]
    emb /= 0.18215
    im = emb.unsqueeze(0)
    im = ldm.decode(im)

    im = TF.to_pil_image(im.squeeze(0).add(1).div(2).clamp(0, 1))

    return [im]
    
css = """
        .gradio-container {
            font-family: 'IBM Plex Sans', sans-serif;
        }
        .gr-button {
            color: white;
            border-color: black;
            background: black;
        }
        input[type='range'] {
            accent-color: black;
        }
        .dark input[type='range'] {
            accent-color: #dfdfdf;
        }
        .container {
            max-width: 730px;
            margin: auto;
            padding-top: 1.5rem;
        }
        #gallery {
            min-height: 22rem;
            margin-bottom: 15px;
            margin-left: auto;
            margin-right: auto;
            border-bottom-right-radius: .5rem !important;
            border-bottom-left-radius: .5rem !important;
        }
        #gallery>div>.h-full {
            min-height: 20rem;
        }
        .details:hover {
            text-decoration: underline;
        }
        .gr-button {
            white-space: nowrap;
        }
        .gr-button:focus {
            border-color: rgb(147 197 253 / var(--tw-border-opacity));
            outline: none;
            box-shadow: var(--tw-ring-offset-shadow), var(--tw-ring-shadow), var(--tw-shadow, 0 0 #0000);
            --tw-border-opacity: 1;
            --tw-ring-offset-shadow: var(--tw-ring-inset) 0 0 0 var(--tw-ring-offset-width) var(--tw-ring-offset-color);
            --tw-ring-shadow: var(--tw-ring-inset) 0 0 0 calc(3px var(--tw-ring-offset-width)) var(--tw-ring-color);
            --tw-ring-color: rgb(191 219 254 / var(--tw-ring-opacity));
            --tw-ring-opacity: .5;
        }
        #advanced-btn {
            font-size: .7rem !important;
            line-height: 19px;
            margin-top: 12px;
            margin-bottom: 12px;
            padding: 2px 8px;
            border-radius: 14px !important;
        }
        #advanced-options, #style-options {
            margin-bottom: 20px;
        }
        .footer {
            margin-bottom: 45px;
            margin-top: 35px;
            text-align: center;
            border-bottom: 1px solid #e5e5e5;
        }
        .footer>p {
            font-size: .8rem;
            display: inline-block;
            padding: 0 10px;
            transform: translateY(10px);
            background: white;
        }
        .dark .footer {
            border-color: #303030;
        }
        .dark .footer>p {
            background: #0b0f19;
        }
        .acknowledgments h4{
            margin: 1.25em 0 .25em 0;
            font-weight: bold;
            font-size: 115%;
        }
"""

block = gr.Blocks(css=css)

examples = [
    [
        'A high tech solarpunk utopia in the Amazon rainforest',
        4,
        45,
        7.5,
        1024,
    ],
    [
        'A pikachu fine dining with a view to the Eiffel Tower',
        4,
        45,
        7,
        1024,
    ],
    [
        'A mecha robot in a favela in expressionist style',
        4,
        45,
        7,
        1024,
    ],
    [
        'an insect robot preparing a delicious meal',
        4,
        45,
        7,
        1024,
    ],
    [
        "A small cabin on top of a snowy mountain in the style of Disney, artstation",
        4,
        45,
        7,
        1024,
    ],
]

with block:
    gr.HTML(
        """
            <div style="text-align: center; max-width: 650px; margin: 0 auto;">
              <div
                style="
                  display: inline-flex;
                  align-items: center;
                  gap: 0.8rem;
                  font-size: 1.75rem;
                "
              >
                <svg
                  width="0.65em"
                  height="0.65em"
                  viewBox="0 0 115 115"
                  fill="none"
                  xmlns="http://www.w3.org/2000/svg"
                >
                  <rect width="23" height="23" fill="white"></rect>
                  <rect y="69" width="23" height="23" fill="white"></rect>
                  <rect x="23" width="23" height="23" fill="#AEAEAE"></rect>
                  <rect x="23" y="69" width="23" height="23" fill="#AEAEAE"></rect>
                  <rect x="46" width="23" height="23" fill="white"></rect>
                  <rect x="46" y="69" width="23" height="23" fill="white"></rect>
                  <rect x="69" width="23" height="23" fill="black"></rect>
                  <rect x="69" y="69" width="23" height="23" fill="black"></rect>
                  <rect x="92" width="23" height="23" fill="#D9D9D9"></rect>
                  <rect x="92" y="69" width="23" height="23" fill="#AEAEAE"></rect>
                  <rect x="115" y="46" width="23" height="23" fill="white"></rect>
                  <rect x="115" y="115" width="23" height="23" fill="white"></rect>
                  <rect x="115" y="69" width="23" height="23" fill="#D9D9D9"></rect>
                  <rect x="92" y="46" width="23" height="23" fill="#AEAEAE"></rect>
                  <rect x="92" y="115" width="23" height="23" fill="#AEAEAE"></rect>
                  <rect x="92" y="69" width="23" height="23" fill="white"></rect>
                  <rect x="69" y="46" width="23" height="23" fill="white"></rect>
                  <rect x="69" y="115" width="23" height="23" fill="white"></rect>
                  <rect x="69" y="69" width="23" height="23" fill="#D9D9D9"></rect>
                  <rect x="46" y="46" width="23" height="23" fill="black"></rect>
                  <rect x="46" y="115" width="23" height="23" fill="black"></rect>
                  <rect x="46" y="69" width="23" height="23" fill="black"></rect>
                  <rect x="23" y="46" width="23" height="23" fill="#D9D9D9"></rect>
                  <rect x="23" y="115" width="23" height="23" fill="#AEAEAE"></rect>
                  <rect x="23" y="69" width="23" height="23" fill="black"></rect>
                </svg>
                <h1 style="font-weight: 900; margin-bottom: 7px;">
                  Classifier Guided Stable Diffusion
                </h1>
              </div>
              <p style="margin-bottom: 10px; font-size: 94%">
                a custom version of stable diffusion with classifier guidance
              </p>
            </div>
        """
    )
    with gr.Group():
        with gr.Box():
            with gr.Row().style(mobile_collapse=False, equal_height=True):
                text = gr.Textbox(
                    label="Enter your prompt",
                    show_label=False,
                    max_lines=1,
                    placeholder="Enter your prompt",
                ).style(
                    border=(True, False, True, True),
                    rounded=(True, False, False, True),
                    container=False,
                )
                btn = gr.Button("Generate image").style(
                    margin=False,
                    rounded=(False, True, True, False),
                )

        gallery = gr.Gallery(
            label="Generated images", show_label=False, elem_id="gallery"
        ).style(grid=[2], height="auto")

        #advanced_button = gr.Button("Advanced options", elem_id="advanced-btn")

        with gr.Row(elem_id="style-options"):
            style = gr.Radio(["none","photo","digital art","anime"], label="Image style")
        with gr.Row(elem_id="advanced-options"):
            #samples = gr.Slider(label="Images", minimum=1, maximum=4, value=4, step=1)
            #steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=45, step=1)
            scale = gr.Slider(
                label="CFG Scale", minimum=0, maximum=50, value=7.5, step=0.1
            )
            classifier_scale = gr.Slider(
                label="Classifier Scale", minimum=0, maximum=1000, value=100, step=1
            )
            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=2147483647,
                step=1,
                randomize=True,
            )

        ex = gr.Examples(examples=examples, fn=infer, inputs=[text, style, scale, classifier_scale, seed], outputs=gallery, cache_examples=True)
        ex.dataset.headers = [""]

        
        text.submit(infer, inputs=[text, style, scale, classifier_scale, seed], outputs=gallery)
        btn.click(infer, inputs=[text, style, scale, classifier_scale, seed], outputs=gallery)

        gr.HTML(
            """
                <div class="footer">
                    <p>Model by <a href="https://huggingface.co./CompVis" style="text-decoration: underline;" target="_blank">CompVis</a> and <a href="https://huggingface.co./stabilityai" style="text-decoration: underline;" target="_blank">Stability AI</a> - Gradio Demo by 🤗 Hugging Face
                    </p>
                </div>
                <div class="acknowledgments">
                    <p><h4>LICENSE</h4>
The model is licensed with a <a href="https://huggingface.co./spaces/CompVis/stable-diffusion-license" style="text-decoration: underline;" target="_blank">CreativeML Open RAIL-M</a> license. The authors claim no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in this license. The license forbids you from sharing any content that violates any laws, produce any harm to a person, disseminate any personal information that would be meant for harm, spread misinformation and target vulnerable groups. For the full list of restrictions please <a href="https://huggingface.co./spaces/CompVis/stable-diffusion-license" target="_blank" style="text-decoration: underline;" target="_blank">read the license</a></p>
                    <p><h4>Biases and content acknowledgment</h4>
Despite how impressive being able to turn text into image is, beware to the fact that this model may output content that reinforces or exacerbates societal biases, as well as realistic faces, pornography and violence. The model was trained on the <a href="https://laion.ai/blog/laion-5b/" style="text-decoration: underline;" target="_blank">LAION-5B dataset</a>, which scraped non-curated image-text-pairs from the internet (the exception being the removal of illegal content) and is meant for research purposes. You can read more in the <a href="https://huggingface.co./CompVis/stable-diffusion-v1-4" style="text-decoration: underline;" target="_blank">model card</a></p>
               </div>
           """
        )

block.queue(max_size=25).launch()