Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1176,10 +1176,16 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
             
         # Concatenate the image embeddings
        ## clip_image_embeds = torch.mean(torch.stack(image_prompt_embeds_list), dim=0)
-        clip_image_embeds = torch.cat(image_prompt_embeds_list, dim=0).mean(dim=0) #.unsqueeze(0)
-        bs_embed, seq_len = clip_image_embeds.shape
-        clip_image_embeds = clip_image_embeds.view(bs_embed, seq_len)  # Simplified reshape
 
+        
+       # clip_image_embeds = torch.cat(image_prompt_embeds_list, dim=0).mean(dim=0) #.unsqueeze(0)
+        #bs_embed, seq_len = clip_image_embeds.shape
+        #clip_image_embeds = clip_image_embeds.view(bs_embed, seq_len)  # Simplified reshape
+
+        # experimental way
+        clip_image_embeds = torch.cat([torch.zeros_like(image_prompt_embeds_list), image_prompt_embeds_list], dim=0).mean(dim=0)
+
+        
         # 4. Prepare timesteps
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)