Update fake_video.py

fake_video.py

@@ -35,7 +35,6 @@ torch.load = _patched_load
 print("🚀 Initializing Advanced Voting System with OpenCLIP & Claude...")
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
-# API Key for Claude via OpenRouter securely fetched from HF Secrets
 OPENROUTER_API_KEY = os.environ.get("OPENROUTER_API_KEY")
 
 def clear_memory():
@@ -45,7 +44,7 @@ def clear_memory():
 
 clear_memory()
 
-# --- 2. OPEN CLIP PROMPTS (15 Real, 15 Fake) ---
+# --- 2. OPEN CLIP PROMPTS ---
 REAL_PROMPTS = [
     "authentic high quality photo", "natural skin texture", "real human face", 
     "genuine unaltered photograph", "natural facial imperfections", "realistic human portrait", 
@@ -65,24 +64,19 @@ FAKE_PROMPTS = [
 # --- 3. LOAD MODELS ---
 print("Loading Models (HF ViT + OpenCLIP)...")
 
-# A. HuggingFace Deepfake ViT
 model_id = "dima806/deepfake_vs_real_image_detection"
 processor = AutoImageProcessor.from_pretrained(model_id, token=os.environ.get("HF_TOKEN"))
 hf_model = AutoModelForImageClassification.from_pretrained(model_id, token=os.environ.get("HF_TOKEN")).to(DEVICE)
 if DEVICE == "cuda": hf_model = hf_model.half()
 hf_model.eval()
 
-# B. OpenCLIP
 clip_model, _, clip_transform = open_clip.create_model_and_transforms('ViT-B-32', pretrained='laion2b_s34b_b79k')
 clip_model.to(DEVICE).eval()
 tokenizer = open_clip.get_tokenizer('ViT-B-32')
 
-# Pre-compute text tokens for CLIP to save time during loops
 with torch.no_grad():
     real_tokens = tokenizer(REAL_PROMPTS).to(DEVICE)
     fake_tokens = tokenizer(FAKE_PROMPTS).to(DEVICE)
-    
-    # Text Embeddings & Normalization
     real_text_embs = F.normalize(clip_model.encode_text(real_tokens), dim=-1)
     fake_text_embs = F.normalize(clip_model.encode_text(fake_tokens), dim=-1)
 
@@ -90,7 +84,10 @@ print("✅ All Models Loaded")
 
 # --- 4. UTILS & MATH ---
 def extract_faces_retina(frame_img):
+    # 🚀 SPEED OPTIMIZATION: Shrink image before passing to heavy RetinaFace model
+    frame_img.thumbnail((512, 512), Image.Resampling.LANCZOS)
     frame_np = np.array(frame_img)
+    
     try: faces = RetinaFace.detect_faces(frame_np)
     except: return frame_img
 
@@ -123,17 +120,12 @@ def extract_10_frames(video_path):
     cap.release()
     return frames
 
-# Mathematical Embeddings & Cosine Similarity
 def get_clip_similarity(face_img):
     img_tensor = clip_transform(face_img).unsqueeze(0).to(DEVICE)
     with torch.no_grad():
-        # 1. Extract & Normalize Image Embedding
         img_emb = F.normalize(clip_model.encode_image(img_tensor), dim=-1)
-        
-        # 2. Cosine Similarity (Dot Product of Normalized Vectors)
         sim_real = (img_emb @ real_text_embs.T).mean().item()
         sim_fake = (img_emb @ fake_text_embs.T).mean().item()
-        
     return sim_real, sim_fake
 
 # --- 5. CLAUDE 3.5 SONNET ---
@@ -150,7 +142,7 @@ def get_claude_reasoning(image, fake_votes, real_votes, sim_real, sim_fake, verd
         f"- Final System Verdict: {verdict}\n\n"
         f"Instruction:\n"
         f"Based on the data and the provided frame image, write EXACTLY 2 lines of reasoning explaining why this video is {verdict}. "
-        f"Use very simple, easy-to-understand English. Do not use difficult technical words or complex jargon. Just explain it simply."
+        f"Use very simple, easy-to-understand English. Do not use difficult technical words or complex jargon."
     )
 
     payload = {
@@ -170,10 +162,12 @@ def get_claude_reasoning(image, fake_votes, real_votes, sim_real, sim_fake, verd
         return "Reasoning unavailable due to API error."
 
 # --- 6. MAIN PIPELINE ---
-def analyze_video(video):
+# 🚀 UI OPTIMIZATION: Added gr.Progress so user sees live loading bar
+def analyze_video(video, progress=gr.Progress()):
     clear_memory()
     if video is None: return None, "No video uploaded.", "Error"
 
+    progress(0, desc="Extracting frames from video...")
     frames = extract_10_frames(video)
     if not frames: return None, "Could not extract frames.", "Error"
 
@@ -186,16 +180,17 @@ def analyze_video(video):
     max_fake_conf = 0
 
     print("🔍 Analyzing 10 frames...")
-    for frame in frames:
+    for i, frame in enumerate(frames):
+        # 🚀 UI OPTIMIZATION: Update loading bar for every frame
+        progress((i + 1) / 10, desc=f"Analyzing Frame {i + 1} of 10...")
+        
         face = extract_faces_retina(frame)
         processed_faces.append(face)
 
-        # A. Semantic Similarity (OpenCLIP)
         sim_r, sim_f = get_clip_similarity(face)
         total_sim_real += sim_r
         total_sim_fake += sim_f
 
-        # B. HF Deepfake Model
         with torch.no_grad():
             inputs = processor(images=face, return_tensors="pt").to(DEVICE)
             if DEVICE == "cuda":
@@ -207,7 +202,6 @@ def analyze_video(video):
             label = hf_model.config.id2label[pred_idx].lower()
             confidence = torch.softmax(logits, dim=1)[0][pred_idx].item() * 100
 
-            # Frame Voting Logic (> 75% confidence to count as fake)
             if "fake" in label and confidence > 75.0:
                 fake_votes += 1
                 if confidence > max_fake_conf:
@@ -216,18 +210,15 @@ def analyze_video(video):
             else:
                 real_votes += 1
 
-    # Ensure we have an image to show Claude
     if best_face_for_claude is None:
         best_face_for_claude = processed_faces[0]
 
-    # --- FINAL DECISION ---
+    progress(0.95, desc="Writing Forensic Report with Claude...")
     avg_sim_real = total_sim_real / 10
     avg_sim_fake = total_sim_fake / 10
     is_forged = fake_votes >= 6
     final_verdict = "FORGED" if is_forged else "REAL"
 
-    # Fetch 2-line reasoning from Claude
-    print("🧠 Fetching Claude Reasoning...")
     claude_reasoning = get_claude_reasoning(
         best_face_for_claude, fake_votes, real_votes, avg_sim_real, avg_sim_fake, final_verdict
     )
@@ -245,6 +236,4 @@ def analyze_video(video):
         f"> {claude_reasoning}"
     )
 
-    return best_face_for_claude, summary, verdict_text
-
-# Removed the standalone UI launch block as requested
+    return best_face_for_claude, summary, verdict_text