Unloading

app.py

@@ -1259,12 +1259,15 @@ def worker_video_end(input_video, end_frame, prompts, n_prompt, seed, batch, res
         stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'CLIP Vision encoding ...'))))
 
         if not high_vram:
-            load_model_as_complete(image_encoder, vae, target_device=gpu)
+            load_model_as_complete(image_encoder, target_device=gpu)
 
         image_encoder_output = hf_clip_vision_encode(input_image_np, feature_extractor, image_encoder)
 
         # 20250507 pftq: Process end frame if provided
         if end_frame is not None:
+            if not high_vram:
+                load_model_as_complete(vae, target_device=gpu)
+
             stream.output_queue.push(('progress', (None, '', make_progress_bar_html(0, 'End frame encoding ...'))))
             end_latent, end_clip_embedding = image_encode(
                 end_frame, target_width=width, target_height=height, vae=vae,
@@ -1841,8 +1844,8 @@ with block:
 
             with gr.Accordion("Debug", open=False):
                 input_image_debug = gr.Image(type="numpy", label="Image Debug", height=320)
-                end_image_debug = gr.Image(type="numpy", label="End Image Debug", height=320)
                 input_video_debug = gr.Video(sources='upload', label="Input Video Debug", height=320)
+                end_image_debug = gr.Image(type="numpy", label="End Image Debug", height=320)
                 prompt_debug = gr.Textbox(label="Prompt Debug", value='')
                 total_second_length_debug = gr.Slider(label="Additional Video Length to Generate (seconds) Debug", minimum=1, maximum=120, value=1, step=0.1)