Update generator.py

generator.py

@@ -763,14 +763,16 @@ class RetroArtConverter:
         pipe_kwargs["generator"] = generator
         
         # --- START FIX: Correct Compel batching and slicing ---
+        # This fixes the "93 vs 77" error
         if self.use_compel and self.compel is not None:
             try:
                 print("Encoding prompts with Compel...")
                 
-                # Pass both prompts as a list to be batched - Compel will pad them to match
+                # Pass both prompts as a list to be batched
                 conditioning_batch, pooled_batch = self.compel([prompt, negative_prompt])
-            
+
                 # Unpack the batch results using slicing
+                # [0:1] and [1:2] keeps the batch dimension, which is required
                 pipe_kwargs["prompt_embeds"] = conditioning_batch[0:1]
                 pipe_kwargs["pooled_prompt_embeds"] = pooled_batch[0:1]
                 pipe_kwargs["negative_prompt_embeds"] = conditioning_batch[1:2]
@@ -829,28 +831,36 @@ class RetroArtConverter:
                         
                         print(f"  - Face embedding: {face_emb_tensor.shape} -> {face_proj_embeds.shape}, Scale: {boosted_scale:.2f}")
                         
+                        # --- START FIX: Pad negative embeds to match face embeds ---
+                        # This fixes the "109 vs 93" error
                         if 'prompt_embeds' in pipe_kwargs:
                             original_embeds = pipe_kwargs['prompt_embeds']
                             
-                            if original_embeds.shape[0] > 1: # Handle CFG
-                                face_proj_embeds = torch.cat([torch.zeros_like(face_proj_embeds), face_proj_embeds], dim=0)
+                            # Note: The old CFG check (if original_embeds.shape[0] > 1:) is removed
+                            # as our Compel logic already provides separate cond/uncond embeds.
                             
                             combined_embeds = torch.cat([original_embeds, face_proj_embeds], dim=1)
                             pipe_kwargs['prompt_embeds'] = combined_embeds
-                            
+
                             # CRITICAL: Pad negative_prompt_embeds by the same amount
                             if 'negative_prompt_embeds' in pipe_kwargs:
                                 negative_embeds = pipe_kwargs['negative_prompt_embeds']
                                 neg_padding = torch.zeros(
-                                    negative_embeds.shape[0],
-                                    face_proj_embeds.shape[1],
-                                    negative_embeds.shape[2],
+                                    (
+                                        negative_embeds.shape[0],      # 1
+                                        face_proj_embeds.shape[1], # 16
+                                        negative_embeds.shape[2],      # 2048
+                                    ),
                                     device=negative_embeds.device,
                                     dtype=negative_embeds.dtype
                                 )
                                 pipe_kwargs['negative_prompt_embeds'] = torch.cat([negative_embeds, neg_padding], dim=1)
+                                print(f"  [OK] Negative prompt padded to match: {pipe_kwargs['negative_prompt_embeds'].shape}")
+                            
+                            print(f"  [OK] Face embeddings concatenated successfully! Prompt: {combined_embeds.shape}")
                         else:
                             print(f"  [WARNING] Can't concatenate - no prompt_embeds (use Compel)")
+                        # --- END FIX ---
                 
                 elif has_detected_faces:
                     print("  Face detected but IP-Adapter/embeddings unavailable, using keypoints only")