Update pipeline.py

pipeline.py

@@ -1494,26 +1494,34 @@ class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdap
 
                     # sum the noise predictions for the unconditional and text conditioned noise
                     start_guidance_time = time.time()
-                    
-                    # if do_classifier_free_guidance:
-                    #     noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-
-                    #     # add the ending frames from noise_pred_uncond to the start of the noise_pred_uncond_sum
-                    #     noise_pred_uncond_sum[:, :,current_context_indexes, :, :] += noise_pred_uncond
-                    #     noise_pred_text_sum[:, :,current_context_indexes, :, :] += noise_pred_text
-                    #     #increase the counter for the ending frames
-                    #     latent_counter[current_context_indexes] += 1
-                        
                         
                     if do_classifier_free_guidance:
+                        # # Split tensor along its first dimension
+                        # noise_pred_uncond, noise_pred_text = torch.chunk(noise_pred, 2, dim=0)
+
+                        # # Efficient in-place addition using advanced indexing
+                        # noise_pred_uncond_sum[..., current_context_indexes, :, :] += noise_pred_uncond
+                        # noise_pred_text_sum[..., current_context_indexes, :, :] += noise_pred_text
+
+                        # # Efficient in-place increment for latent_counter
+                        # latent_counter[current_context_indexes] += 1
+                        
+                        
                         # Split tensor along its first dimension
                         noise_pred_uncond, noise_pred_text = torch.chunk(noise_pred, 2, dim=0)
 
-                        # Efficient in-place addition using advanced indexing
-                        noise_pred_uncond_sum[..., current_context_indexes, :, :] += noise_pred_uncond
-                        noise_pred_text_sum[..., current_context_indexes, :, :] += noise_pred_text
+                        # Reshape or expand noise_pred_uncond and noise_pred_text to match the dimensions of the sum tensors
+                        # This step depends on the dimensions of your tensors and how they need to align for the operation
+                        expanded_noise_pred_uncond = noise_pred_uncond.expand_as(noise_pred_uncond_sum[..., current_context_indexes, :, :])
+                        expanded_noise_pred_text = noise_pred_text.expand_as(noise_pred_text_sum[..., current_context_indexes, :, :])
+
+                        # Perform batch addition
+                        noise_pred_uncond_sum[..., current_context_indexes, :, :] += expanded_noise_pred_uncond
+                        noise_pred_text_sum[..., current_context_indexes, :, :] += expanded_noise_pred_text
 
-                        # Efficient in-place increment for latent_counter
+                        # Batch increment for latent_counter
+                        # Here, you need to ensure that the addition is done correctly depending on how latent_counter is structured
+                        # If latent_counter is a tensor, you might use something like:
                         latent_counter[current_context_indexes] += 1
                         
                     print("guidance time", time.time() - start_guidance_time)