README.md

metadata

title: CC-4DGS
emoji: 🎞️
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 5.50.0
python_version: '3.10'
app_file: app.py
pinned: false
license: other
short_description: Storage-efficient dynamic Gaussian splatting viewer

CC-4DGS viewer (ZeroGPU)

Interactive renderer for CC-4DGS: Computational Deformation and Point-Cloud Compression for Storage-Efficient Dynamic Gaussian Splatting (IEEE TVCG 2026, arXiv:2609.02184, code).

Load a trained scene and render it at any time step from the dataset cameras, the N3DV spiral trajectory, or an orbit:

• CDF checkpoints: point_cloud/iteration_<N>/{point_cloud.ply, deformation.pth}
• CDF+CCA checkpoints: point_cloud/iteration_<N>/{compressed/, deformation.pth}
• optional cameras: poses_bounds.npy (N3DV) or COLMAP images.bin + cameras.bin (Technicolor)

Upload a zip of the scene folder or point the app at a folder in a Hugging Face model/dataset repo. Set the Space variables CC4DGS_HUB_REPO, CC4DGS_HUB_SUBFOLDER and CC4DGS_HUB_REPO_TYPE to pre-fill the Hub fields; set the HF_TOKEN secret for private repos.

Only rendering runs on the GPU (@spaces.GPU); checkpoint decoding, CCA decompression and camera setup run on CPU. The rasterizer and simple-knn CUDA extensions are installed at start-up from wheels/ (sm_80/86/89/90), or compiled from submodules/ when the wheels do not match the installed torch.

License

The CC-4DGS code is Apache-2.0. The Inria Gaussian Splatting rasterizer and simple-knn are restricted to non-commercial research and evaluation use; see LICENSE.md and THIRD_PARTY_NOTICES.md.