URA is now 0.3.1 with torch 2.9

app.py

@@ -3,7 +3,6 @@
 from __future__ import annotations
 
 import sys
-import threading
 from pathlib import Path
 
 # Allow importing unreflectanything when run from gradio_space (e.g. HF Space with root dir)
@@ -67,29 +66,27 @@ def _get_sample_images():
 # Single model instance; loaded in background at app start or on first inference.
 _cached_ura_model = None
 _cached_device = None
-_model_load_lock = threading.Lock()
 
 
 def _get_model(device: str):
     """Return the pretrained model, loading it once and reusing. Ensures weights exist (downloads if missing)."""
     global _cached_ura_model, _cached_device
     weights_path, config_path = _ensure_weights()
-    with _model_load_lock:
-        if _cached_ura_model is not None and _cached_device == device:
-            return _cached_ura_model
-        from unreflectanything import model
-
-        _cached_ura_model = model(
-            pretrained=True,
-            # weights_path=os.path.join(os.path.dirname(__file__), ".cache", "weights", "full_model_weights.pt"),
-            # config_path=os.path.join(os.path.dirname(__file__), ".cache", "configs", "pretrained_config.yaml"),
-            weights_path=weights_path,
-            config_path=config_path,
-            device=device,
-            verbose=False,
-        )
-        _cached_device = device
+    if _cached_ura_model is not None and _cached_device == device:
         return _cached_ura_model
+    from unreflectanything import model
+
+    _cached_ura_model = model(
+        pretrained=True,
+        # weights_path=os.path.join(os.path.dirname(__file__), ".cache", "weights", "full_model_weights.pt"),
+        # config_path=os.path.join(os.path.dirname(__file__), ".cache", "configs", "pretrained_config.yaml"),
+        weights_path=weights_path,
+        config_path=config_path,
+        device=device,
+        verbose=False,
+    )
+    _cached_device = device
+    return _cached_ura_model
 
 
 def build_ui():
@@ -97,7 +94,8 @@ def build_ui():
 
     device = "cuda" if torch.cuda.is_available() else "cpu"
     # Start loading the model in the background so it is ready (or nearly ready) by first use.
-    threading.Thread(target=_get_model, args=(device,), daemon=True).start()
+    print(f"Initializing model on {device}...")
+    _get_model(device)
 
     def run_inference(image: np.ndarray | None) -> np.ndarray | None:
         """Run reflection removal using the cached model. Returns RGB numpy [H,W,3] in 0–255 or None."""
@@ -211,7 +209,6 @@ def _launch_allowed_paths():
 
 if __name__ == "__main__":
     demo.launch(
-        share=True,
         allowed_paths=_launch_allowed_paths(),
         theme=gr.themes.Soft(primary_hue="orange", secondary_hue="blue"),
     )