[start_latent, image_encoder_last_hidden_state] = get_start_latent(input_image, height, width, vae, gpu, image_encoder, high_vram)

app.py

@@ -381,12 +381,12 @@ def worker(input_image, prompts, n_prompt, seed, resolution, total_second_length
             
             if not high_vram:
                 load_model_as_complete(image_encoder, target_device=gpu)
+
+            image_encoder_last_hidden_state = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder).last_hidden_state
                 
-            return start_latent
+            return [start_latent, image_encoder_last_hidden_state]
             
-        start_latent = get_start_latent(input_image, height, width, vae, gpu, image_encoder, high_vram)
-
-        image_encoder_last_hidden_state = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder).last_hidden_state
+        [start_latent, image_encoder_last_hidden_state] = get_start_latent(input_image, height, width, vae, gpu, image_encoder, high_vram)
 
         # Dtype
 
@@ -602,12 +602,12 @@ def worker_last_frame(input_image, prompts, n_prompt, seed, resolution, total_se
             
             if not high_vram:
                 load_model_as_complete(image_encoder, target_device=gpu)
+
+            image_encoder_last_hidden_state = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder).last_hidden_state
                 
-            return start_latent
+            return [start_latent, image_encoder_last_hidden_state]
             
-        start_latent = get_start_latent(input_image, height, width, vae, gpu, image_encoder, high_vram)
-
-        image_encoder_last_hidden_state = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder).last_hidden_state
+        [start_latent, image_encoder_last_hidden_state] = get_start_latent(input_image, height, width, vae, gpu, image_encoder, high_vram)
 
         # Dtype