app.py

import gradio as gr
from PIL import Image  

import torch
import re
import os
import requests

from customization import customize_vae_decoder
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel, DDIMScheduler, EulerDiscreteScheduler
from torchvision import transforms
from attribution import MappingNetwork

import math
from typing import List
from PIL import Image, ImageChops
import numpy as np
import torch


PRETRAINED_MODEL_NAME_OR_PATH = "./checkpoints/"


def get_image_grid(images: List[Image.Image]) -> Image:
    num_images = len(images)
    cols = 3#int(math.ceil(math.sqrt(num_images)))
    rows = 1#int(math.ceil(num_images / cols))
    width, height = images[0].size
    grid_image = Image.new('RGB', (cols * width, rows * height))
    for i, img in enumerate(images):
        x = i % cols
        y = i // cols
        grid_image.paste(img, (x * width, y * height))
    return grid_image


class AttributionModel:
    def __init__(self):
        is_cuda = False
        if torch.cuda.is_available():
            is_cuda = True
        
        scheduler = EulerDiscreteScheduler.from_pretrained('stabilityai/stable-diffusion-2', subfolder="scheduler")
        self.pipe = StableDiffusionPipeline.from_pretrained('stabilityai/stable-diffusion-2', scheduler=scheduler)#, safety_checker=None, torch_dtype=torch.float16)
        if is_cuda:
            self.pipe = self.pipe.to("cuda")
        self.resize_transform = transforms.Resize(512, interpolation=transforms.InterpolationMode.BILINEAR)
        self.vae = AutoencoderKL.from_pretrained(
            'stabilityai/stable-diffusion-2', subfolder="vae"
        )
        self.vae = customize_vae_decoder(self.vae, 128, "deqkv", "all", False, 1.0)

        self.mapping_network = MappingNetwork(32, 0, 128, None, num_layers=2, w_avg_beta=None, normalization = False)
        
        from torchvision.models import resnet50, ResNet50_Weights
        self.decoding_network = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
        self.decoding_network.fc = torch.nn.Linear(2048,32)
        
        self.vae.decoder.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'vae_decoder.pth')))
        self.mapping_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'mapping_network.pth')))
        self.decoding_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'decoding_network.pth')))

        if is_cuda:
            self.vae = self.vae.to("cuda")
            self.mapping_network = self.mapping_network.to("cuda")
            self.decoding_network = self.decoding_network.to("cuda")

        self.test_norm = transforms.Compose(
            [
                transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
            ]
        )

    def infer(self, prompt, negative, steps, guidance_scale):
        with torch.no_grad():
            out_latents = self.pipe([prompt], negative_prompt=[negative], output_type="latent", num_inference_steps=steps, guidance_scale=guidance_scale).images
        image_attr = self.inference_with_attribution(out_latents)
        image_attr_pil = self.pipe.numpy_to_pil(image_attr[0])

        image_org = self.inference_without_attribution(out_latents)
        image_org_pil = self.pipe.numpy_to_pil(image_org[0])

        # image_diff_pil = self.pipe.numpy_to_pil(image_attr[0] - image_org[0])
        diff_factor = 5
        image_diff_pil = ImageChops.difference(image_org_pil[0], image_attr_pil[0]).convert("RGB", (diff_factor,0,0,0,0,diff_factor,0,0,0,0,diff_factor,0))

        return image_org_pil[0], image_attr_pil[0], image_diff_pil

    def inference_without_attribution(self, latents):
        latents = 1 / 0.18215 * latents
        with torch.no_grad():
            image = self.pipe.vae.decode(latents).sample
        image = image.clamp(-1,1)
        image = (image / 2 + 0.5).clamp(0, 1)
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        return image

    def get_phis(self, phi_dimension, batch_size ,eps = 1e-8):
        phi_length = phi_dimension
        b = batch_size
        phi = torch.empty(b,phi_length).uniform_(0,1)
        return torch.bernoulli(phi) + eps


    def inference_with_attribution(self, latents, key=None):
        if key==None:
            key = self.get_phis(32, 1)

        latents = 1 / 0.18215 * latents
        with torch.no_grad():
            image = self.vae.decode(latents, self.mapping_network(key.cuda())).sample
        image = image.clamp(-1,1)
        image = (image / 2 + 0.5).clamp(0, 1)
        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
        return image

    def postprocess(self, image):
        image = self.resize_transform(image)
        return image

    def detect_key(self, image):
        reconstructed_keys = self.decoding_network(self.test_norm((image / 2 + 0.5).clamp(0, 1)))
        return reconstructed_keys


attribution_model = AttributionModel()
def get_images(prompt, negative, steps, guidence_scale):
    x1, x2, x3 = attribution_model.infer(prompt, negative, steps, guidence_scale)
    return [x1, x2, x3]


image_examples = [
    ["A pikachu fine dining with a view to the Eiffel Tower", "low quality", 50, 10],
    ["A mecha robot in a favela in expressionist style", "low quality, 3d, photorealistic", 50, 10]
]

with gr.Blocks() as demo:
    gr.Markdown(
        """<h1 style="text-align: center;"><b>WOUAF:
Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models</b> <br> <a href="https://wouaf.vercel.app">Project Page</a></h1>""")

    with gr.Row(elem_id="prompt-container").style(mobile_collapse=False, equal_height=True):
        with gr.Column():
            text = gr.Textbox(
                label="Enter your prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                elem_id="prompt-text-input",
            ).style(
                border=(True, False, True, True),
                rounded=(True, False, False, True),
                container=False,
            )
            negative = gr.Textbox(
                label="Enter your negative prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter a negative prompt",
                elem_id="negative-prompt-text-input",
            ).style(
                border=(True, False, True, True),
                rounded=(True, False, False, True),
                container=False,
            )

    with gr.Row():
        steps = gr.Slider(label="Steps", minimum=45, maximum=55, value=50, step=1)
        guidance_scale = gr.Slider(
            label="Guidance Scale", minimum=0, maximum=10, value=7.5, step=0.1
        )

    with gr.Row():
        btn = gr.Button(value="Generate Image", full_width=False)

    with gr.Row():
        im_2 = gr.Image(type="pil", label="without attribution")
        im_3 = gr.Image(type="pil", label="**with** attribution")
        im_4 = gr.Image(type="pil", label="pixel-wise difference multiplied by 5")

    
    btn.click(get_images, inputs=[text, negative, steps, guidance_scale], outputs=[im_2, im_3, im_4])

    gr.Examples(
        examples=image_examples,
        inputs=[text, negative, steps, guidance_scale],
        outputs=[im_2, im_3, im_4],
        fn=get_images,
        cache_examples=True,
    )

    gr.HTML(
            """
                <div class="footer">
                    <p>Pre-trained model by <a href="https://huggingface.co/stabilityai" style="text-decoration: underline;" target="_blank">StabilityAI</a>
                    </p>
                    <p>
                    Fine-tuned by authors for research purpose.
                    </p>
                </div>
           """
        )
    with gr.Accordion(label="Ethics & Privacy", open=False):
        gr.HTML(
            """<div class="acknowledgments">
                <p><h4>Privacy</h4>
We do not collect any images or key data. This demo is designed with sole purpose of fun and reducing misuse of AI.
                <p><h4>Biases and content acknowledgment</h4>
This model will have the same biases as Stable Diffusion V2.1               </div>
            """
        )

if __name__ == "__main__":
    demo.launch()