Drops the dependency on the private hysts/StyleGAN3 repo by
downloading from NVIDIA NGC, the official upstream distribution
point. dnnlib.util.open_url handles HTTP fetch + on-disk caching
under ~/.cache/dnnlib/downloads/.
All 12 pkls are pre-fetched at module load so model switching never
blocks on the network. NGC pkls are full bundles (G/D/G_ema/...) so
the total download is ~3GB vs ~720MB on the HF Hub mirror; cold
start measured locally at ~78s, well within the Space startup
timeout. Warm cache start ~4s.
- Migrate to uv (pyproject.toml + uv.lock + .python-version)
- Bump gradio 4.36.1 -> 6.13.0, torch 2.0.1 -> 2.9.1, drop torchvision
- Inline model.py into app.py
- Patch torch.utils.cpp_extension.load to register the loaded module
in sys.modules. StyleGAN3's torch_utils/custom_ops.py calls
importlib.import_module(name) after cpp_extension.load(name=...);
torch>=2.9 no longer auto-registers, so the follow-up import fails
with ModuleNotFoundError even when compilation succeeds.
- Add randomize seed toggle
