Upload generator.py

generator.py

@@ -223,22 +223,8 @@ class RetroArtConverter:
             # Detect faces
             has_detected_faces = False
             face_kps_image = None
+            face_embeddings = None
             face_bbox_original = None
-    
-            # === FIX: START ===
-            # The InstantID pipeline requires image_embeds. We must provide  a dummy tensor if no face is detected.
-    
-            # Get the expected embedding dimension from the pipeline
-            try:
-                # Get dim from the loaded model
-                image_emb_dim = self.pipe.image_proj_model_in_features
-            except AttributeError:
-                print("[WARN] Cannot find image_proj_model_in_features, defaulting to 512")
-                image_emb_dim = 512 # Default dim for InstantID
-
-            # Initialize with dummy embeddings (zeros)
-            face_embeddings = np.zeros(image_emb_dim, dtype=np.float32) # Or dtype=dtype ???
-            # === FIX: END ===
             
             if self.face_detection_enabled and self.face_app is not None:
                 try:
@@ -323,53 +309,52 @@ class RetroArtConverter:
             else:
                 pipe_kwargs["prompt"] = prompt
                 pipe_kwargs["negative_prompt"] = negative_prompt
-
-
+            
             # Configure ControlNets + IP-Adapter (SIMPLIFIED!)
-            # *** ALWAYS PASS EMBEDDINGS ***
-            # The pipeline always needs image_embeds.
-            # This will be the REAL embeddings or the DUMMY (zero) embeddings.
-            pipe_kwargs["image_embeds"] = face_embeddings
-    
             if has_detected_faces and face_kps_image is not None:
                 print("Using InstantID (keypoints + embeddings) + Depth ControlNets")
-    
+                
                 # Control images: [face keypoints, depth map]
                 pipe_kwargs["control_image"] = [face_kps_image, depth_image]
-    
+                
                 # Conditioning scales: [identity, depth]
                 pipe_kwargs["controlnet_conditioning_scale"] = [
                     identity_control_scale,
                     depth_control_scale
                 ]
-    
-                # Control guidance timing
+                
+                # Control guidance timing (when each ControlNet is active)
+                # [start, start] - both active from beginning
+                # [end, end] - both active until end
                 pipe_kwargs["control_guidance_start"] = [0.0, 0.0]
                 pipe_kwargs["control_guidance_end"] = [1.0, 1.0]
-    
-                # Control IP-Adapter strength for REAL embeddings
-                pipe_kwargs["ip_adapter_scale"] = identity_preservation
-    
-                print(f"  - Face embeddings shape: {face_embeddings.shape}")
-                print(f"  - IP-Adapter scale: {identity_preservation}")
-    
+                
+                # IP-Adapter face embeddings (SIMPLE - pipeline handles everything!)
+                if face_embeddings is not None:
+                    print(f"Adding face embeddings for IP-Adapter...")
+                    
+                    # Just pass the embeddings - pipeline does the rest!
+                    pipe_kwargs["image_embeds"] = face_embeddings
+                    
+                    # Control IP-Adapter strength
+                    pipe_kwargs["ip_adapter_scale"] = identity_preservation
+                    
+                    print(f"  - Face embeddings shape: {face_embeddings.shape}")
+                    print(f"  - IP-Adapter scale: {identity_preservation}")
+                    print(f"  [OK] Face embeddings configured")
+                else:
+                    print("  [WARNING] No face embeddings - using keypoints only")
+            
             else:
                 print("No faces detected - using Depth ControlNet only")
-    
+                
                 # Use depth for both ControlNet slots (identity scale = 0)
                 pipe_kwargs["control_image"] = [depth_image, depth_image]
                 pipe_kwargs["controlnet_conditioning_scale"] = [0.0, depth_control_scale]
-    
+                
                 # Control guidance timing for both slots
                 pipe_kwargs["control_guidance_start"] = [0.0, 0.0]
                 pipe_kwargs["control_guidance_end"] = [1.0, 1.0]
-    
-                # *** SET IP-ADAPTER SCALE TO 0 ***
-                # We passed dummy embeddings, so set scale to 0
-                pipe_kwargs["ip_adapter_scale"] = 0.0
-    
-                print(f"  - Dummy embeddings shape: {face_embeddings.shape}")
-                print(f"  - IP-Adapter scale: 0.0")
             
             # Generate
             print(f"Generating: Steps={num_inference_steps}, CFG={guidance_scale}, Strength={strength}")