import os import sys import torch import torchvision.transforms as transforms import torchvision.transforms.functional as TF from PIL import Image # Add UniversalFakeDetect to path BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) UFD_DIR = os.path.join(BASE_DIR, "UniversalFakeDetect") if UFD_DIR not in sys.path: sys.path.insert(0, UFD_DIR) from models import get_model # Constants from validate.py MEAN_CLIP = [0.48145466, 0.4578275, 0.40821073] STD_CLIP = [0.26862954, 0.26130258, 0.27577711] # 1. Pipeline A: validate.py (RealFakeDataset in validate.py) transform_validate_py = transforms.Compose([ transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean=MEAN_CLIP, std=STD_CLIP) ]) # 2. Pipeline B: datasets.py (data_loader in train.py / val_opt with loadSize=256, cropSize=224) transform_train_val = transforms.Compose([ transforms.Resize(256, interpolation=transforms.InterpolationMode.BICUBIC), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean=MEAN_CLIP, std=STD_CLIP) ]) def main(): print("=================================================================") print("TASK 5: Comparing Official Inference Implementations vs Current App") print("=================================================================\n") device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"Using device: {device}") # Load Model exactly as official validate.py print("Loading CLIP:ViT-L/14 model...") model = get_model("CLIP:ViT-L/14") ckpt_path = os.path.join(UFD_DIR, "pretrained_weights", "fc_weights.pth") state_dict = torch.load(ckpt_path, map_location="cpu") model.fc.load_state_dict(state_dict) model.eval() model.to(device) print(f"Model and weights loaded from {ckpt_path}\n") # Pipeline C: clip.load's preprocess (used in current app.py) transform_app_py = model.preprocess test_images = [ ("Known real photograph", r"C:\Users\user\.gemini\antigravity-ide\brain\3b54ff09-c09f-4fe6-9276-f4cca9fbc6d7\scratch\real_phone_photo.png", "REAL"), ("Known AI-generated image", r"C:\Users\user\.gemini\antigravity-ide\brain\3b54ff09-c09f-4fe6-9276-f4cca9fbc6d7\ai_cyberpunk_city_1789660863192.jpg", "FAKE"), ("Known manipulated image", r"C:\Users\user\.gemini\antigravity-ide\brain\3b54ff09-c09f-4fe6-9276-f4cca9fbc6d7\scratch\manipulated_image.png", "EVALUATE") ] pipelines = [ ("Current app.py (model.preprocess: Resize(224, bicubic) -> CenterCrop(224))", transform_app_py), ("validate.py (Direct CenterCrop(224) without resizing)", transform_validate_py), ("train.py/datasets.py (Resize(256, bicubic) -> CenterCrop(224))", transform_train_val) ] results = [] for img_name, img_path, expected_label in test_images: print(f"-----------------------------------------------------------------") print(f"Test Image: {img_name}") print(f"Path: {img_path}") print(f"Expected / Label: {expected_label}") with Image.open(img_path) as raw_img: img_rgb = raw_img.convert("RGB") print(f"Original size: {img_rgb.size} ({img_rgb.width}x{img_rgb.height})") for pipe_name, pipe_transform in pipelines: # Apply transform tensor = pipe_transform(img_rgb).unsqueeze(0).to(device) with torch.inference_mode(): feat = model.model.encode_image(tensor) logit = model.fc(feat.float()).item() prob_fake = torch.sigmoid(torch.tensor(logit)).item() pred_label = "FAKE" if prob_fake >= 0.50 else "REAL" conf = prob_fake if pred_label == "FAKE" else (1.0 - prob_fake) print(f" [{pipe_name}]") print(f" Raw Logit: {logit:+.5f}") print(f" Fake Prob: {prob_fake:.6f} ({prob_fake*100:.2f}%)") print(f" Prediction: {pred_label} (Confidence: {conf*100:.2f}%)") results.append({ "image": img_name, "pipeline": pipe_name, "logit": logit, "prob_fake": prob_fake, "prediction": pred_label, "confidence": conf }) print() print("=================================================================") print("Summary Analysis:") print("=================================================================") for r in results: print(f"{r['image'][:25]:<25} | {r['pipeline'][:35]:<35} | logit: {r['logit']:+8.4f} | prob: {r['prob_fake']*100:6.2f}% | pred: {r['prediction']}") if __name__ == "__main__": main()