Update app.py

app.py

@@ -29,8 +29,8 @@ def detect(image, confidence_threshold=0.5):
             logits = outputs.logits
             probabilities = torch.softmax(logits, dim=1)[0]
         
-        confidence_real = probabilities[0].item() * 100
-        confidence_fake = probabilities[1].item() * 100
+        confidence_real = probabilities[0].item() * 100  # Probability of being Real
+        confidence_fake = probabilities[1].item() * 100  # Probability of being Fake
         
         id2label = model.config.id2label
         predicted_class = torch.argmax(logits, dim=1).item()
@@ -38,17 +38,21 @@ def detect(image, confidence_threshold=0.5):
         threshold_predicted = "Fake" if confidence_fake / 100 >= confidence_threshold else "Real"
         confidence_score = max(confidence_real, confidence_fake)
         
-        # Differentiate outputs (example heuristic)
+        # Use raw probabilities for clarity, with optional fine-tuning for specific categories
+        aigen_likelihood = confidence_fake  # Assuming AI-Generated is synonymous with Fake for simplicity
+        face_manipulation_likelihood = confidence_fake  # You might want to refine this with additional analysis
+        
+        # Format outputs
         overall = f"{confidence_score:.1f}% Confidence ({threshold_predicted})"
-        aigen = f"{confidence_fake * 0.9:.1f}% (AI-Generated Content Likelihood)"  # Arbitrary scaling
-        deepfake = f"{confidence_fake:.1f}% (Face Manipulation Likelihood)"
+        aigen = f"{aigen_likelihood:.1f}% (AI-Generated Content Likelihood)"
+        deepfake = f"{face_manipulation_likelihood:.1f}% (Face Manipulation Likelihood)"
         
         return overall, aigen, deepfake
     
     except Exception as e:
         logger.error(f"Error during analysis: {str(e)}")
         raise gr.Error(f"Analysis error: {str(e)}")
-# Custom CSS (unchanged)
+
 custom_css = """
 .container {
     max-width: 1200px;
@@ -91,6 +95,11 @@ MARKDOWN0 = """
     <h1>DeepFake Detection System</h1>
     <p>Advanced AI-powered analysis for identifying manipulated media<br>
     Powered by prithivMLmods/Deep-Fake-Detector-Model (Updated Jan 2025)<br>
+    <ul>
+        <li><strong>Confidence Score:</strong> Overall probability the image is Real or Fake (based on threshold).</li>
+        <li><strong>AI-Generated Content Likelihood:</strong> Probability the image was fully generated by AI.</li>
+        <li><strong>Face Manipulation Likelihood:</strong> Probability the image contains manipulated faces (e.g., swaps or alterations).</li>
+    </ul>
     Adjust threshold to tune sensitivity; check logs for detailed output</p>
 </div>
 """