Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -1150,10 +1150,14 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
             #clip_image_embeds_1 = self.encode_clip_image_emb(clip_image, device, dtype)
             with torch.inference_mode():
                 clip_image_embeds_1 = self.clip_image_processor(images=clip_image, return_tensors="pt").pixel_values
-                print('clip output size: ', clip_image_embeds_1.shape)
+                print('clip output shape: ', clip_image_embeds_1.shape)
                 clip_image_embeds_1 = clip_image_embeds_1.to(device, dtype=dtype)
                 clip_image_embeds_1 = self.image_encoder(clip_image_embeds_1, output_hidden_states=True).hidden_states[-2]
-            print('encoder output size: ', clip_image_embeds_1.shape)
+                print('encoder output shape: ', clip_image_embeds_1.shape)
+                clip_image_embeds_1 = self.image_proj_model(clip_image_embeds_1)
+            
+            print('projection model output shape: ', clip_image_embeds_1.shape)
+            
             clip_image_embeds_1 = clip_image_embeds_1 * scale_1
             image_prompt_embeds_list.append(clip_image_embeds_1)
         if clip_image_2 != None:
@@ -1163,6 +1167,7 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
                 clip_image_embeds_2 = self.clip_image_processor(images=clip_image_2, return_tensors="pt").pixel_values
                 clip_image_embeds_2 = clip_image_embeds_2.to(device, dtype=dtype)
                 clip_image_embeds_2 = self.image_encoder(clip_image_embeds_2, output_hidden_states=True).hidden_states[-2]
+                clip_image_embeds_2 = self.image_proj_model(clip_image_embeds_2)
             clip_image_embeds_2 = clip_image_embeds_2 * scale_2
             image_prompt_embeds_list.append(clip_image_embeds_2)
         if clip_image_3 != None:
@@ -1172,6 +1177,7 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
                 clip_image_embeds_3 = self.clip_image_processor(images=clip_image_3, return_tensors="pt").pixel_values
                 clip_image_embeds_3 = clip_image_embeds_3.to(device, dtype=dtype)
                 clip_image_embeds_3 = self.image_encoder(clip_image_embeds_3, output_hidden_states=True).hidden_states[-2]
+                clip_image_embeds_3 = self.image_proj_model(clip_image_embeds_3)
             clip_image_embeds_3 = clip_image_embeds_3 * scale_3
             image_prompt_embeds_list.append(clip_image_embeds_3)
         if clip_image_4 != None:
@@ -1181,6 +1187,7 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
                 clip_image_embeds_4 = self.clip_image_processor(images=clip_image_4, return_tensors="pt").pixel_values
                 clip_image_embeds_4 = clip_image_embeds_4.to(device, dtype=dtype)
                 clip_image_embeds_4 = self.image_encoder(clip_image_embeds_4, output_hidden_states=True).hidden_states[-2]
+                clip_image_embeds_4 = self.image_proj_model(clip_image_embeds_4)
             clip_image_embeds_4 = clip_image_embeds_4 * scale_4
             image_prompt_embeds_list.append(clip_image_embeds_4)
         if clip_image_5 != None:
@@ -1190,11 +1197,12 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
                 clip_image_embeds_5 = self.clip_image_processor(images=clip_image_5, return_tensors="pt").pixel_values
                 clip_image_embeds_5 = clip_image_embeds_5.to(device, dtype=dtype)
                 clip_image_embeds_5 = self.image_encoder(clip_image_embeds_5, output_hidden_states=True).hidden_states[-2]
+                clip_image_embeds_5 = self.image_proj_model(clip_image_embeds_5)
             clip_image_embeds_5 = clip_image_embeds_5 * scale_5
             image_prompt_embeds_list.append(clip_image_embeds_5)
             
-        clip_image_embeds_cat_list = torch.cat(image_prompt_embeds_list).mean(dim=0)
-        print('catted embeds list with mean: ',clip_image_embeds_cat_list.shape)
+        clip_image_embeds_cat_list = torch.cat(image_prompt_embeds_list).mean(dim=0).unsqueeze(0)
+        print('catted embeds list with mean and unsqueeze: ',clip_image_embeds_cat_list.shape)
         seq_len, _ = clip_image_embeds_cat_list.shape
         clip_image_embeds_cat_list_repeat = clip_image_embeds_cat_list.repeat(1, 1, 1)
         print('catted embeds repeat: ',clip_image_embeds_cat_list_repeat.shape)