Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1215,24 +1215,28 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
         print('shape 2: ', average_image_embed.shape)
         average_image_embedf = torch.mean(stacked_image_embeds, dim=1).unsqueeze(0)  # Add batch dimension after averaging
         print('shape 2a: ', average_image_embedf.shape)
-        average_image_embede = torch.mean(stacked_image_embeds, dim=1)  # Add batch dimension after averaging
-        print('shape 2b: ', average_image_embede.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)
+        zeros_tensor = torch.zeros_like(average_image_embed)
+        zeros_tensora = average_image_embed.repeat(1, 1, 1)
+        print('shape 3.1: ', clip_image_embedsa.shape)
+        clip_image_embedsa = average_image_embed.repeat(1, 1, 1)
+        print('shape 3.5: ', clip_image_embedsa.shape)
+        clip_image_embedse = torch.cat([zeros_tensora, average_image_embeda], dim=0)
+        print('shape 3.8: ', clip_image_embedse.shape)
         # 4. Concatenate the zeros and the average embedding
         clip_image_embeds = torch.cat([zeros_tensor, average_image_embed], dim=0)
-        clip_image_embeds2 = torch.cat([zeros_tensor, average_image_embede], dim=0)
         print('shape 4: ', clip_image_embeds.shape)
         print('shape 4a: ', clip_image_embeds2.shape)
+        clip_image_embeds2 = torch.cat([zeros_tensor, average_image_embede], dim=0)
         '''
         #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)
-        print('shape 6: ', clip_image_embeds.shape)
-        
+
         clip_image_embedsa = clip_image_embeds.view(bs_embed, 1, -1)
         print('shape 7: ', clip_image_embedsa.shape)
         clip_image_embedsb = clip_image_embeds.view(seq_len, -1)