from io import BytesIO
import requests
import gradio as gr
import requests
import torch
from tqdm import tqdm
from PIL import Image, ImageOps
from diffusers import StableDiffusionInpaintPipeline
from torchvision.transforms import ToPILImage
from utils import preprocess, prepare_mask_and_masked_image, recover_image, resize_and_crop
gr.close_all()
topil = ToPILImage()
pipe_inpaint = StableDiffusionInpaintPipeline.from_pretrained(
"runwayml/stable-diffusion-inpainting",
revision="fp16",
torch_dtype=torch.float16,
)
pipe_inpaint = pipe_inpaint.to("cuda")
## Good params for editing that we used all over the paper --> decent quality and speed
GUIDANCE_SCALE = 7.5
NUM_INFERENCE_STEPS = 100
DEFAULT_SEED = 1234
def pgd(X, targets, model, criterion, eps=0.1, step_size=0.015, iters=40, clamp_min=0, clamp_max=1, mask=None):
X_adv = X.clone().detach() + (torch.rand(*X.shape)*2*eps-eps).cuda()
pbar = tqdm(range(iters))
for i in pbar:
actual_step_size = step_size - (step_size - step_size / 100) / iters * i
X_adv.requires_grad_(True)
loss = (model(X_adv).latent_dist.mean - targets).norm()
pbar.set_description(f"Loss {loss.item():.5f} | step size: {actual_step_size:.4}")
grad, = torch.autograd.grad(loss, [X_adv])
X_adv = X_adv - grad.detach().sign() * actual_step_size
X_adv = torch.minimum(torch.maximum(X_adv, X - eps), X + eps)
X_adv.data = torch.clamp(X_adv, min=clamp_min, max=clamp_max)
X_adv.grad = None
if mask is not None:
X_adv.data *= mask
return X_adv
def get_target():
target_url = 'https://www.rtings.com/images/test-materials/2015/204_Gray_Uniformity.png'
response = requests.get(target_url)
target_image = Image.open(BytesIO(response.content)).convert("RGB")
target_image = target_image.resize((512, 512))
return target_image
def immunize_fn(init_image, mask_image):
with torch.autocast('cuda'):
mask, X = prepare_mask_and_masked_image(init_image, mask_image)
X = X.half().cuda()
mask = mask.half().cuda()
targets = pipe_inpaint.vae.encode(preprocess(get_target()).half().cuda()).latent_dist.mean
adv_X = pgd(X,
targets = targets,
model=pipe_inpaint.vae.encode,
criterion=torch.nn.MSELoss(),
clamp_min=-1,
clamp_max=1,
eps=0.12,
step_size=0.01,
iters=200,
mask=1-mask
)
adv_X = (adv_X / 2 + 0.5).clamp(0, 1)
adv_image = topil(adv_X[0]).convert("RGB")
adv_image = recover_image(adv_image, init_image, mask_image, background=True)
return adv_image
def run(image, prompt, seed, guidance_scale, num_inference_steps, immunize=False):
if seed == '':
seed = DEFAULT_SEED
else:
seed = int(seed)
torch.manual_seed(seed)
init_image = Image.fromarray(image['image'])
init_image = resize_and_crop(init_image, (512,512))
mask_image = ImageOps.invert(Image.fromarray(image['mask']).convert('RGB'))
mask_image = resize_and_crop(mask_image, init_image.size)
if immunize:
immunized_image = immunize_fn(init_image, mask_image)
image_edited = pipe_inpaint(prompt=prompt,
image=init_image if not immunize else immunized_image,
mask_image=mask_image,
height = init_image.size[0],
width = init_image.size[1],
eta=1,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps,
).images[0]
image_edited = recover_image(image_edited, init_image, mask_image)
if immunize:
return [(immunized_image, 'Immunized Image'), (image_edited, 'Edited After Immunization')]
else:
return [(image_edited, 'Edited Image (Without Immunization)')]
description='''Demo of our paper:
**Raising the Cost of Malicious AI-Powered Image Editing**
*[Hadi Salman](https://twitter.com/hadisalmanX), [Alaa Khaddaj](https://twitter.com/Alaa_Khaddaj), [Guillaume Leclerc](https://twitter.com/gpoleclerc), [Andrew Ilyas](https://twitter.com/andrew_ilyas), [Aleksander Madry](https://twitter.com/aleks_madry)*
MIT [Paper](https://arxiv.org/abs/2302.06588)
[Blog post](https://gradientscience.org/photoguard/)
[![](https://badgen.net/badge/icon/GitHub?icon=github&label)](https://github.com/MadryLab/photoguard)
Below you can test our (encoder attack) immunization method for making images resistant to manipulation by Stable Diffusion. This immunization process forces the model to perform unrealistic edits.
**See Section 5 in our paper for a discussion of the intended use cases for (as well as limitations of) this tool.**
'''
examples_list = [
['./images/hadi_and_trevor.jpg', 'man attending a wedding', '329357', GUIDANCE_SCALE, NUM_INFERENCE_STEPS],
['./images/trevor_2.jpg', 'two men in prison', '329357', GUIDANCE_SCALE, NUM_INFERENCE_STEPS],
['./images/elon_2.jpg', 'man in a metro station', '214213', GUIDANCE_SCALE, NUM_INFERENCE_STEPS],
]
with gr.Blocks() as demo:
gr.HTML(value="""
For faster inference without waiting in queue, run this demo locally (instruction in our Github repo), or duplicate this space and upgrade to GPU in settings.