Update generator.py

generator.py

@@ -767,17 +767,16 @@ class RetroArtConverter:
             try:
                 print("Encoding prompts with Compel...")
                 
-                # Pass both prompts as a list to be batched
+                # Pass both prompts as a list to be batched - Compel will pad them to match
                 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]
                 pipe_kwargs["negative_pooled_prompt_embeds"] = pooled_batch[1:2]
                 
-                print("[OK] Using Compel-encoded prompts")
+                print(f"[OK] Compel encoded - Prompt: {pipe_kwargs['prompt_embeds'].shape}, Negative: {pipe_kwargs['negative_prompt_embeds'].shape}")
             except Exception as e:
                 print(f"Compel encoding failed, using standard prompts: {e}")
                 traceback.print_exc()
@@ -838,7 +837,18 @@ class RetroArtConverter:
                             
                             combined_embeds = torch.cat([original_embeds, face_proj_embeds], dim=1)
                             pipe_kwargs['prompt_embeds'] = combined_embeds
-                            print(f"  [OK] Face embeddings concatenated successfully! New shape: {combined_embeds.shape}")
+                            
+                            # 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],
+                                    device=negative_embeds.device,
+                                    dtype=negative_embeds.dtype
+                                )
+                                pipe_kwargs['negative_prompt_embeds'] = torch.cat([negative_embeds, neg_padding], dim=1)
                         else:
                             print(f"  [WARNING] Can't concatenate - no prompt_embeds (use Compel)")