Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1192,14 +1192,13 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
             clip_image_embeds_5 = self.image_encoder(clip_image_embeds_5, output_hidden_states=True).hidden_states[-2]
             clip_image_embeds_5 = clip_image_embeds_5 * scale_5
             image_prompt_embeds_list.append(clip_image_embeds_5)
-            
-        clip_image_embeds_cat = torch.cat(image_prompt_embeds_list, dim=0) 
-        clip_image_embeds_mean = clip_image_embeds_cat.mean(dim=0, keepdim=True)
-        #clip_image_embeds_cat_list = torch.cat(image_prompt_embeds_list).mean(dim=1,keepdim=True)
-        print('catted embeds list with mean: ', clip_image_embeds_mean.shape)
-        zeros_tensor = torch.zeros_like(clip_image_embeds_mean)
+
+            #cat, but not mean
+        clip_image_embeds_cat = torch.cat(image_prompt_embeds_list) 
+        print('catted embeds list without mean: ', clip_image_embeds_cat.shape)
+        zeros_tensor = torch.zeros_like(clip_image_embeds_cat)
         print('zeros: ',zeros_tensor.shape)
-        clip_image_embeds = torch.cat([zeros_tensor, clip_image_embeds_mean], dim=0)
+        clip_image_embeds = torch.cat([zeros_tensor, clip_image_embeds_mean], dim=1)
         print('embeds shape: ', clip_image_embeds.shape)
         
         # 4. Prepare timesteps