Update ip_adapter/ip_adapter.py

ip_adapter/ip_adapter.py

@@ -109,6 +109,7 @@ class IPAdapter:
         clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
         clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.bfloat16)).image_embeds
         image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        print('image_proj_model shape:',image_prompt_embeds)
         uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
         return image_prompt_embeds, uncond_image_prompt_embeds
     
@@ -238,6 +239,7 @@ class IPAdapterXL(IPAdapter):
         print('Using primary image.')
 
         image_prompt_embeds_1, uncond_image_prompt_embeds_1 = self.get_image_embeds(pil_image_1)
+        image_prompt_embeds_1 = image_prompt_embeds_1 * scale_1
         image_prompt_embeds_list.append(image_prompt_embeds_1)
         uncond_image_prompt_embeds_list.append(uncond_image_prompt_embeds_1)
         
@@ -267,10 +269,12 @@ class IPAdapterXL(IPAdapter):
             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)