Upload app.py with huggingface_hub

app.py

@@ -50,6 +50,13 @@ snapshot_download(
 snapshot_download(repo_id=SDV2_REPO, local_dir=CKPT_DIR,
                   allow_patterns=["wan_causal_dmd_v2v/*"])
 
+# Pre-fetch the TAEHV tiny-VAE decoder weights (fast streaming decode).
+_TAEHV_PATH = os.path.join(CKPT_DIR, "taew2_1.pth")
+if not os.path.exists(_TAEHV_PATH):
+    import urllib.request
+    urllib.request.urlretrieve(
+        "https://github.com/madebyollin/taehv/raw/main/taew2_1.pth", _TAEHV_PATH)
+
 device = torch.device("cuda")
 
 # StreamDiffusionV2 single-GPU streaming pipeline (rolling KV + sink tokens are
@@ -60,10 +67,10 @@ stream = StreamDiffusionV2Pipeline(
     device=device,
     height=HEIGHT,
     width=WIDTH,
-    step=2,
+    step=1,            # fewer denoising stages -> shallower pipeline -> lower lag
     noise_scale=NOISE_SCALE,
     model_type="T2V-1.3B",
-    use_taehv=False,
+    use_taehv=True,    # tiny-VAE decode -> much faster per-chunk -> lower lag
 )
 PM = stream.pipeline_manager
 CHUNK = PM.base_chunk_size * PM.pipeline.num_frame_per_block   # 4 px frames / chunk