Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1154,9 +1154,6 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
                 clip_image_embeds_1 = clip_image_embeds_1.to(device, dtype=dtype)
                 clip_image_embeds_1 = self.image_encoder(clip_image_embeds_1, output_hidden_states=True).hidden_states[-2]
                 print('encoder output shape: ', clip_image_embeds_1.shape)
-            
-            print('projection model output shape: ', clip_image_embeds_1.shape)
-            
             clip_image_embeds_1 = clip_image_embeds_1 * scale_1
             image_prompt_embeds_list.append(clip_image_embeds_1)
         if clip_image_2 != None:
@@ -1199,11 +1196,11 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
         clip_image_embeds_cat_list = torch.cat(image_prompt_embeds_list).mean(dim=1)
         #clip_image_embeds_cat_list = torch.cat(image_prompt_embeds_list, dim=1)
         print('catted embeds list with mean: ',clip_image_embeds_cat_list.shape)
-        #seq_len, _ = clip_image_embeds_cat_list.shape
-        #clip_image_embeds_cat_list_repeat = clip_image_embeds_cat_list.repeat(1, 1, 1)
+        seq_len, _ = clip_image_embeds_cat_list.shape
+        clip_image_embeds_cat_list_repeat = clip_image_embeds_cat_list.repeat(1, 1, 1)
         clip_image_embeds_view = clip_image_embeds_cat_list #.unsqueeze(0) # Added unsqueeze here instead
-        #print('catted embeds repeat: ',clip_image_embeds_view.shape)
-        #clip_image_embeds_view = clip_image_embeds_view.view(1, seq_len, -1)
+        print('catted embeds repeat: ',clip_image_embeds_view.shape)
+        clip_image_embeds_view = clip_image_embeds_view.view(1, seq_len, -1)
         print('catted viewed: ',clip_image_embeds_view.shape)
         zeros_tensor = torch.zeros_like(clip_image_embeds_view)
         #zeros_tensor = torch.zeros_like(clip_image_embeds_view[:, :image_prompt_embeds_list[0].shape[1], :]) # Make zeros tensor match the sequence length of a single image embedding