Update pipeline_stable_diffusion_3_ipa.py

pipeline_stable_diffusion_3_ipa.py

@@ -965,7 +965,17 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
 
         # ipa
         clip_image=None,
+        clip_image_2=None,
+        clip_image_3=None,
+        clip_image_4=None,
+        clip_image_5=None,
+        text_scale=1.0,
         ipadapter_scale=1.0,
+        scale_1=1.0,
+        scale_2=1.0,
+        scale_3=1.0,
+        scale_4=1.0,
+        scale_5=1.0,
     ):
         r"""
         Function invoked when calling the pipeline for generation.
@@ -1126,10 +1136,42 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
             pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
-
+            
+        prompt_embeds = prompt_embeds * text_scale
+        
+        image_prompt_embeds_list = []
+        
         # 3. prepare clip emb
         clip_image = clip_image.resize((max(clip_image.size), max(clip_image.size)))
-        clip_image_embeds = self.encode_clip_image_emb(clip_image, device, dtype)
+        clip_image_embeds_1 = self.encode_clip_image_emb(clip_image, device, dtype)
+        image_prompt_embeds_list.append(clip_image_embeds_1)
+
+        if clip_image_2 != None:
+            print('Using secondary image.')
+            clip_image_2 = clip_image_2.resize((max(clip_image.size), max(clip_image.size)))
+            image_prompt_embeds_2 = self.encode_clip_image_emb(clip_image, device, dtype)
+            image_prompt_embeds_2 = image_prompt_embeds_2 * scale_2
+            image_prompt_embeds_list.append(image_prompt_embeds_2)
+        if clip_image_3 != None:
+            print('Using tertiary image.')
+            clip_image_3 = clip_image_3.resize((max(clip_image.size), max(clip_image.size)))
+            image_prompt_embeds_3 = self.encode_clip_image_emb(clip_image, device, dtype)
+            image_prompt_embeds_3 = image_prompt_embeds_3 * scale_3
+            image_prompt_embeds_list.append(image_prompt_embeds_3)
+        if clip_image_4 != None:
+            print('Using quaternary image.')
+            clip_image_4 = clip_image_4.resize((max(clip_image.size), max(clip_image.size)))
+            image_prompt_embeds_4 = self.encode_clip_image_emb(clip_image, device, dtype)
+            image_prompt_embeds_4 = image_prompt_embeds_4 * scale_4
+            image_prompt_embeds_list.append(image_prompt_embeds_4)
+        if clip_image_5 != None:
+            print('Using quinary image.')
+            clip_image_5 = clip_image_5.resize((max(clip_image.size), max(clip_image.size)))
+            image_prompt_embeds_5 = self.encode_clip_image_emb(clip_image, device, dtype)
+            image_prompt_embeds_5 = image_prompt_embeds_5 * scale_5
+            image_prompt_embeds_list.append(image_prompt_embeds_5)
+
+        clip_image_embeds = torch.cat(image_prompt_embeds_list).mean(dim=0).unsqueeze(0)
 
         # 4. Prepare timesteps
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)