More optimization

app.py

@@ -995,20 +995,21 @@ def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed,
                 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_latent = image_encode(
                 end_frame, target_width=width, target_height=height, vae=vae,
                 image_encoder=image_encoder, feature_extractor=feature_extractor, device=gpu
-            )[:2]
+            )[:1]
             end_frame = None
             end_latent = end_latent.to(dtype=torch.float32, device=cpu)
         else:
-            end_latent = end_clip_embedding = None
+            end_latent = None
 
         # Clean GPU
         if not high_vram:
             unload_complete_models(image_encoder, vae)
 
         image_encoder_last_hidden_state = image_encoder_output.last_hidden_state
+        image_encoder_output = None
 
         # Dtype
         image_encoder_last_hidden_state = image_encoder_last_hidden_state.to(transformer.dtype)
@@ -1180,6 +1181,13 @@ def worker_video(input_video, end_frame, end_stillness, prompts, n_prompt, seed,
                     clean_latent_4x_indices=clean_latent_4x_indices,
                     callback=callback,
                 )
+                clean_latents = None
+                clean_latents_2x = None
+                clean_latents_4x = None
+                latent_indices = None
+                clean_latent_indices = None
+                clean_latent_2x_indices = None
+                clean_latent_4x_indices = None
 
                 total_generated_latent_frames += int(generated_latents.shape[2])
                 history_latents = torch.cat([history_latents, generated_latents.to(history_latents)], dim=2)