Update pipeline.py

pipeline.py

@@ -1539,10 +1539,15 @@ class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdap
                 start_id = batch * output_batch_size
                 end_id = min((batch + 1) * output_batch_size, num_frames)
                 video_tensor = self.decode_latents(latents[:, :, start_id:end_id, :, :])
+                # if we had more then one prompt, we need to offset the video frames back by number of inference steps
+                if len(prompt_embeds_list) > 1:
+                    # wrap the first n number of frames to the end of the video to fix the offseting from the context scheduler
+                    video_tensor = torch.cat((video_tensor[:, :, num_inference_steps:, :, :], video_tensor[:, :, :num_inference_steps, :, :]), dim=2)
                 video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)
                 for f_id, frame in enumerate(video[0]):
                     frame.save(frame_format.format(start_id + f_id))
             return output_path
+        
             
         # Post-processing
         video_tensor = self.decode_latents(latents)