Update app.py

app.py

@@ -288,7 +288,8 @@ class LTX23DistilledA2VPipeline:
         )
 
         torch.cuda.synchronize()
-        cleanup_memory()
+        del video_state       # release Stage 1 latent only, keep encoder/transformer on GPU
+        torch.cuda.empty_cache()  # lightweight, no sync
 
         # ── Upscaling ──
         upscaled_video_latent = upsample_video(
@@ -324,9 +325,6 @@ class LTX23DistilledA2VPipeline:
         )
 
         torch.cuda.synchronize()
-        del transformer
-        del video_encoder
-        cleanup_memory()
 
         # ── Decode both video and audio ──
         decoded_video = vae_decode_video(