Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1206,21 +1206,24 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
             
         # 1. Stack the image embeddings
         stacked_image_embeds = torch.stack(image_prompt_embeds_list)
-
+        print('shape 1: ', stacked_image_embeds.shape)
         # 2. Calculate the mean of the stacked embeddings
         average_image_embed = torch.mean(stacked_image_embeds, dim=0) #.unsqueeze(0)  # Add batch dimension after averaging
-
+        print('shape 2: ', average_image_embed.shape)
         # 3. Create a tensor of zeros with the same shape as the averaged embedding
         zeros_tensor = torch.zeros_like(average_image_embed)
-
+        print('shape 3: ', zeros_tensor.shape)
         # 4. Concatenate the zeros and the average embedding
         clip_image_embeds = torch.cat([zeros_tensor, average_image_embed], dim=0)
+        print('shape 4: ', clip_image_embeds.shape)
+        clip_image_embeds = clip_image_embeds.unsqueeze(0) # Add a dimension at the beginning so now you have [1, 2*seq_len_img, embed_dim_img]
+        print('shape 5: ', clip_image_embeds.shape)
 
         bs_embed, seq_len, _ = clip_image_embeds.shape
         clip_image_embeds = clip_image_embeds.repeat(1, 1, 1)
-        clip_image_embeds = clip_image_embeds.unsqueeze(0) # Add a dimension at the beginning so now you have [1, 2*seq_len_img, embed_dim_img]
-        clip_image_embeds = clip_image_embeds.view(2, 1, -1)
-            
+        print('shape 6: ', clip_image_embeds.shape)
+        clip_image_embeds = clip_image_embeds.view(bs_embed, 1, -1)
+        print('shape 7: ', clip_image_embeds.shape)
         #clip_image_embeds = torch.cat([torch.stack(image_prompt_embeds_list)], dim=0).mean(dim=0)
         
         # 4. Prepare timesteps