Update app.py

app.py

@@ -675,10 +675,30 @@ def generate_image(
         pipe.scheduler = scheduler_cls.from_config(pipe.scheduler.config, **add_kwargs)
         print(f"  ✓ Scheduler: {scheduler_class_name}")
 
-        # Apply style and process prompts
+        # Apply style and process prompts (AFTER face detection so we can use gender info)
         if not prompt:
             prompt = "a person"
         
+        # Add explicit gender to prompt if detected (InstantID works better with explicit gender)
+        if detected_gender is not None:
+            prompt_lower = prompt.lower()
+            # Only add gender if not already in prompt
+            if "man" not in prompt_lower and "woman" not in prompt_lower and "male" not in prompt_lower and "female" not in prompt_lower and "person" not in prompt_lower:
+                if detected_gender == 0:  # Female
+                    prompt = f"a woman, {prompt}"
+                    print(f"  ✓ Added 'a woman' to prompt for gender preservation")
+                elif detected_gender == 1:  # Male
+                    prompt = f"a man, {prompt}"
+                    print(f"  ✓ Added 'a man' to prompt for gender preservation")
+            elif "person" in prompt_lower:
+                # Replace "person" with specific gender
+                if detected_gender == 0:  # Female
+                    prompt = prompt.replace("person", "woman").replace("Person", "Woman")
+                    print(f"  ✓ Replaced 'person' with 'woman' in prompt")
+                elif detected_gender == 1:  # Male
+                    prompt = prompt.replace("person", "man").replace("Person", "Man")
+                    print(f"  ✓ Replaced 'person' with 'man' in prompt")
+        
         # Warn if prompt contains physical feature descriptions that might override identity
         physical_keywords = ["hair", "blonde", "brown hair", "black hair", "red hair", "beard", "mustache", 
                             "wearing", "shirt", "jacket", "suit", "blazer", "tie", "glasses"]
@@ -692,9 +712,9 @@ def generate_image(
         if detected_gender is not None:
             # Add opposite gender terms to negative prompt
             if detected_gender == 0:  # Female
-                gender_negative_terms += ", man, male, masculine"
+                gender_negative_terms += ", man, male, masculine, boy"
             elif detected_gender == 1:  # Male
-                gender_negative_terms += ", woman, female, feminine"
+                gender_negative_terms += ", woman, female, feminine, girl"
             print(f"  ✓ Gender preservation enabled in negative prompt")
         
         # Add gender preservation terms to negative prompt
@@ -717,15 +737,18 @@ def generate_image(
             # Fallback: try to convert
             face_emb = torch.tensor(face_emb_raw, device=device, dtype=dtype)
         
-        # Normalize face embedding (L2 normalization for better identity preservation)
+        # Ensure proper shape for InstantID (should be [1, 512] or [512])
         if len(face_emb.shape) == 1:
             face_emb = face_emb.unsqueeze(0)  # Add batch dimension: [1, 512]
+        elif len(face_emb.shape) == 2 and face_emb.shape[0] > 1:
+            # If multiple faces, take the first one
+            face_emb = face_emb[0:1]
         
-        # L2 normalize the embedding to unit length (standard for face embeddings)
-        face_emb = torch.nn.functional.normalize(face_emb, p=2, dim=1)
+        # DO NOT normalize - InstantID expects raw embeddings from InsightFace
+        # Normalization breaks the identity preservation
         
         print(f"  Face embedding final shape: {face_emb.shape}, dtype: {face_emb.dtype}, device: {face_emb.device}")
-        print(f"  Face embedding norm: {torch.norm(face_emb, p=2, dim=1).item():.4f} (should be ~1.0)")
+        print(f"  Face embedding range: [{face_emb.min().item():.4f}, {face_emb.max().item():.4f}]")
         
         face_kps = draw_kps(convert_from_cv2_to_image(face_image_cv2), face_info["kps"])
         print(f"  Face keypoints image size: {face_kps.size}")
@@ -1201,19 +1224,19 @@ with gr.Blocks() as demo:
                     identitynet_strength_ratio = gr.Slider(
                         label="Face Similarity",
                         minimum=0.5,
-                        maximum=1.2,
+                        maximum=1.5,
                         step=0.05,
-                        value=1.0,
-                        info="How closely the headshot resembles your photo (higher = more similar)",
+                        value=1.2,
+                        info="How closely the headshot resembles your photo (higher = more similar, recommended: 1.0-1.2)",
                     )
 
                     adapter_strength_ratio = gr.Slider(
                         label="Face Detail Strength",
                         minimum=0.3,
-                        maximum=1.2,
+                        maximum=1.5,
                         step=0.05,
-                        value=0.85,
-                        info="Strength of face features preservation (higher = better identity match)",
+                        value=1.0,
+                        info="Strength of face features preservation (higher = better identity match, recommended: 0.8-1.0)",
                     )
 
                     enable_LCM = gr.Checkbox(