import gradio as gr import os import torch import numpy as np device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') from diffusers import DiffusionPipeline, AutoencoderKL import torchvision.transforms as transforms from copy import deepcopy from collections import OrderedDict import requests import json from PIL import Image, ImageEnhance import base64 import io import random import math class BZHStableSignatureDemo(object): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self.pipe = DiffusionPipeline.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16").to("cuda") # disable invisible-watermark self.pipe.watermark = None # save the original VAE decoders = OrderedDict([("no watermark", self.pipe.vae)]) # load the patched VAEs for name in ("weak", "medium", "strong", "extreme"): vae = AutoencoderKL.from_pretrained(f"imatag/stable-signature-bzh-sdxl-vae-{name}", torch_dtype=torch.float16).to("cuda") decoders[name] = vae self.decoders = decoders def generate(self, mode, seed, prompt): generator = torch.Generator(device=device) torch.manual_seed(seed) # load the patched VAE vae = self.decoders[mode] self.pipe.vae = vae output = self.pipe(prompt, num_inference_steps=4, guidance_scale=0.0, output_type="pil") return output.images[0] def attack_detect(self, img, jpeg_compression, downscale, crop, saturation): img = img.convert("RGB") # attack if downscale != 1: size = img.size size = (int(size[0] / downscale), int(size[1] / downscale)) img = img.resize(size, Image.Resampling.LANCZOS) if crop != 0: width, height = img.size area = width * height log_rmin = math.log(0.5) log_rmax = math.log(2.0) for _ in range(10): target_area = area * (1 - crop) aspect_ratio = math.exp(random.random() * (log_rmax - log_rmin) + log_rmin) w = int(round(math.sqrt(target_area * aspect_ratio))) h = int(round(math.sqrt(target_area / aspect_ratio))) if 0 < w <= width and 0 < h <= height: top = random.randint(0, height - h + 1) left = random.randint(0, width - w + 1) img = img.crop((left, top, left+w, top+h)) break converter = ImageEnhance.Color(img) img = converter.enhance(saturation) # send to detection API and apply JPEG compression attack mf = io.BytesIO() img.save(mf, format='JPEG', quality=jpeg_compression) # includes JPEG attack b64 = base64.b64encode(mf.getvalue()) data = { 'image': b64.decode('utf8') } headers = {} api_key = os.getenv('BZH_API_KEY') if api_key: headers['x-api-key'] = api_key response = requests.post('https://bzh.imatag.com/bzh/api/v1.0/detect', json=data, headers=headers) response.raise_for_status() data = response.json() pvalue = data['p-value'] mf.seek(0) img0 = Image.open(mf) # reload to show JPEG attack result = "No watermark detected." rpv = 10**int(math.log10(pvalue)) if pvalue < 1e-3: result = "Watermark detected with low confidence (p-value<%.0e)" % rpv if pvalue < 1e-9: result = "Watermark detected with high confidence (p-value<%.0e)" % rpv return (img0, result) def interface(): prompt = "sailing ship in storm by Rembrandt" backend = BZHStableSignatureDemo() decoders = list(backend.decoders.keys()) with gr.Blocks() as demo: gr.Markdown("""# Watermarked SDXL-Turbo demo This demo brought to you by [IMATAG](https://www.imatag.com/) presents watermarking of images generated via [StableDiffusion XL Turbo](https://huggingface.co/stabilityai/sdxl-turbo). Using the method presented in [StableSignature](https://ai.meta.com/blog/stable-signature-watermarking-generative-ai/), the VAE decoder of StableDiffusion is fine-tuned to produce images including a specific invisible watermark. We combined this method with a demo version of [IMATAG](https://www.imatag.com/)'s in-house decoder. The watermarking system operates in zero-bit mode for improved robustness.""") with gr.Row(): inp = gr.Textbox(label="Prompt", value=prompt) seed = gr.Number(label="Seed", precision=0) mode = gr.Dropdown(choices=decoders, label="Watermark strength", value="medium") with gr.Row(): btn1 = gr.Button("Generate") with gr.Row(): watermarked_image = gr.Image(type="pil", width=512, height=512) with gr.Column(): gr.Markdown("""With these controls you may alter the generated image before detection. You may also upload your own edited image instead.""") downscale = gr.Slider(1, 3, value=1, step=0.1, label="Downscale ratio") crop = gr.Slider(0, 0.9, value=0, step=0.01, label="Random crop ratio") saturation = gr.Slider(0, 2, value=1, step=0.1, label="Color saturation") jpeg_compression = gr.Slider(value=100, step=5, label="JPEG quality") btn2 = gr.Button("Modify & Detect") with gr.Row(): attacked_image = gr.Image(type="pil", width=256) detection_label = gr.Label(label="Detection info") btn1.click(fn=backend.generate, inputs=[mode, seed, inp], outputs=[watermarked_image], api_name="generate") btn2.click(fn=backend.attack_detect, inputs=[watermarked_image, jpeg_compression, downscale, crop, saturation], outputs=[attacked_image, detection_label], api_name="detect") return demo if __name__ == '__main__': demo = interface() demo.launch()