Update ip_adapter/ip_adapter.py

ip_adapter/ip_adapter.py

@@ -270,16 +270,9 @@ class IPAdapterXL(IPAdapter):
             image_prompt_embeds_list.append(image_prompt_embeds_5)
             uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds_5)
 
-        image_prompt_embeds = torch.cat(image_prompt_embeds_list).mean(dim=0).unsqueeze(0)
-        print('catted embeds list with mean and unsqueeze shape: ',image_prompt_embeds.shape)
-        bs_embed, seq_len, _ = image_prompt_embeds.shape
-        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
-        print('catted embeds repeat: ',image_prompt_embeds.shape)
-        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
-        print('viewed embeds: ',image_prompt_embeds.shape)
-        uncond_image_prompt_embeds = torch.cat(uncond_image_prompt_embeds_list).mean(dim=0).unsqueeze(0)  
-        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
-        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        image_prompt_embeds = torch.cat(image_prompt_embeds_list).mean(dim=0)
+        image_prompt_embeds = image_prompt_embeds.mean(dim=0, keepdim=True)
+        uncond_image_prompt_embeds = torch.cat(uncond_image_prompt_embeds_list).mean(dim=0)
 
         with torch.inference_mode():
             prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = self.pipe.encode_prompt(