Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1204,7 +1204,16 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
         #clip_image_embeds = torch.cat([torch.zeros_like(torch.stack(image_prompt_embeds_list)), torch.stack(image_prompt_embeds_list)], dim=0).mean(dim=0)
         # FAILS clip_image_embeds = torch.cat(torch.stack(image_prompt_embeds_list), dim=0).mean(dim=0)
         # FAILS TIMESTEPS clip_image_embeds = torch.cat(image_prompt_embeds_list, dim=0).mean(dim=0)
+        clip_image_embeds_cat_list = torch.cat(image_prompt_embeds_list).mean(dim=0)
+        print('catted with mean shape 1: ', stacked_image_embeds.shape)
+        clip_image_embeds_stack_list = torch.stack(image_prompt_embeds_list).mean(dim=0)
+        print('stacked with mean shape 1: ', stacked_image_embeds.shape)
+        clip_image_embeds_cat_list = torch.cat(image_prompt_embeds_list)
+        print('catted without mean shape 1: ', stacked_image_embeds.shape)
+        clip_image_embeds_stack_list = torch.stack(image_prompt_embeds_list)
+        print('stacked without mean shape 1: ', stacked_image_embeds.shape)
 
+        
         # 1. Stack the image embeddings
         stacked_image_embeds = torch.cat(image_prompt_embeds_list, dim=1)
         print('shape 1: ', stacked_image_embeds.shape)
@@ -1214,18 +1223,20 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
         # 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_tensora = zeros_tensor.repeat(1, 1, 1)
-        print('shape 3.1: ', zeros_tensora.shape)
-        clip_image_embedsa = average_image_embed.repeat(1, 1, 1)
-        print('shape 3.5: ', clip_image_embedsa.shape)
+        zeros_tensor_repeat = zeros_tensor.repeat(1, 1, 1)
+        print('shape 3.1: ', zeros_tensor_repeat.shape)
+        clip_image_embeds_repeat = average_image_embed.repeat(1, 1, 1)
+        print('shape 3.5: ', clip_image_embeds_repeat.shape)
         # 4. Concatenate the zeros and the average embedding
-        clip_image_embeds2 = torch.cat([zeros_tensor, average_image_embed], dim=0)
-        clip_image_embeds3 = clip_image_embeds2.repeat(1, 1, 1)
-        print('shape 4: ', clip_image_embeds2.shape)
-        clip_image_embeds = torch.cat([zeros_tensora, clip_image_embedsa], dim=0)
-        print('shape 4a: ', clip_image_embeds.shape)
-        clip_image_embeds = torch.cat([zeros_tensora, clip_image_embedsa], dim=0)
-        print('shape 4b: ', clip_image_embeds3.shape)
+        clip_image_embeds_cat = torch.cat([zeros_tensor, average_image_embed], dim=0)
+        print('shape 4: ', clip_image_embeds_cat.shape)
+        clip_image_embeds_cat_repeat = clip_image_embeds_cat.repeat(1, 1, 1)
+        print('shape 4.1: ', clip_image_embeds_cat_repeat.shape)
+        clip_image_embeds_repeat_cat = torch.cat([zeros_tensor_repeat, clip_image_embeds_repeat], dim=0)
+        print('shape 4a: ', clip_image_embeds_repeat_cat.shape)
+        clip_image_embeds_repeat_cat_1 = torch.cat([zeros_tensor_repeat, clip_image_embeds_repeat], dim=1)
+        print('shape 4b: ', clip_image_embeds_repeat_cat_1.shape)
+        clip_image_embeds = clip_image_embeds_repeat_cat
         '''
         #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)