import os import sys import zipfile import requests import pickle import torch from PIL import Image import numpy as np import torchvision.transforms as transforms import cv2 from timm import create_model import gradio as gr # ------------------------------- # تحميل كود StyleGAN2 # ------------------------------- if not os.path.exists("stylegan2-ada-pytorch"): os.system("git clone https://github.com/NVlabs/stylegan2-ada-pytorch.git") sys.path.append("stylegan2-ada-pytorch") # ------------------------------- # تحميل ملفات النماذج المضغوطة # ------------------------------- stylegan2_url = "https://huggingface.co/lam107/stylegan2-models/resolve/main/stylegan2-ffhq-512x512.zip" swin_url = "https://huggingface.co/lam107/swin_classifier/resolve/main/best_model_final.pth.zip" if not os.path.exists("stylegan2-ffhq-512x512.pkl"): print("Downloading StyleGAN2 model...") r = requests.get(stylegan2_url) with open("stylegan2.zip", "wb") as f: f.write(r.content) with zipfile.ZipFile("stylegan2.zip", 'r') as zip_ref: zip_ref.extractall(".") if not os.path.exists("best_model_final.pth"): print("Downloading Swin model...") r = requests.get(swin_url) with open("swin.zip", "wb") as f: f.write(r.content) with zipfile.ZipFile("swin.zip", 'r') as zip_ref: zip_ref.extractall(".") # ------------------------------- # تحميل StyleGAN2 Discriminator # ------------------------------- device = torch.device("cuda" if torch.cuda.is_available() else "cpu") with open("stylegan2-ffhq-512x512.pkl", "rb") as f: data = pickle.load(f) D = data['D'].to(device) D.eval() # ------------------------------- # تحميل Swin Transformer # ------------------------------- class_names = ["real", "fake"] model = create_model('swin_large_patch4_window12_384', pretrained=False, num_classes=len(class_names)) model.load_state_dict(torch.load("best_model_final.pth", map_location=device)) model = model.to(device) model.eval() # ------------------------------- # التدرجات + التصنيف # ------------------------------- def compute_gradients(model, pil_image): transform = transforms.Compose([ transforms.Resize((512, 512)), transforms.ToTensor(), transforms.Normalize([0.5], [0.5]) ]) image = transform(pil_image.convert("RGB")).unsqueeze(0).to(device).requires_grad_(True) c = torch.zeros([image.shape[0], model.c_dim], device=device) if hasattr(model, "c_dim") and model.c_dim > 0 else None output = model(image, c) loss = -output.mean() loss.backward() gradients = image.grad.detach().cpu().numpy() gradients = np.abs(gradients).mean(axis=1)[0] gradients = (gradients - gradients.min()) / (gradients.max() - gradients.min() + 1e-8) gradients = (gradients * 255).astype(np.uint8) grad_color = cv2.applyColorMap(gradients, cv2.COLORMAP_JET) grad_rgb = cv2.cvtColor(grad_color, cv2.COLOR_BGR2RGB) return Image.fromarray(grad_rgb) def swin_classifier(pil_image): transform = transforms.Compose([ transforms.Resize((384, 384)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) image = transform(pil_image).unsqueeze(0).to(device) with torch.no_grad(): output = model(image) probabilities = torch.softmax(output, dim=1).squeeze().cpu().numpy() return {"Real": float(probabilities[0]), "Fake": float(probabilities[1])} def process_image(pil_img): grad_pil = compute_gradients(D, pil_img).resize((384, 384)) prediction = swin_classifier(grad_pil) return grad_pil, prediction # ------------------------------- # Gradio واجهة # ------------------------------- with gr.Blocks() as demo: gr.Markdown("**💜 Mukhbir: AI Generated Image Detection**") with gr.Row(): with gr.Column(): image_input = gr.Image(type="pil", label="Upload Image") clear_button = gr.Button("Clear") submit_button = gr.Button("Submit", variant="primary") with gr.Column(): output_image = gr.Image(label="Gradient Image", type="pil") output_label = gr.Label(label="Prediction") submit_button.click(fn=process_image, inputs=image_input, outputs=[output_image, output_label]) clear_button.click(fn=lambda: (None, None, None), inputs=None, outputs=[image_input, output_image, output_label]) demo.launch(server_name="0.0.0.0", server_port=7860,ssr_mode=False)