--- title: README emoji: 🫧 colorFrom: indigo colorTo: purple sdk: static pinned: false --- # LibGS **[ffgs](https://github.com/libgs/ffgs)** — standard inference for feed-forward 3D Gaussian Splatting models. Posed (or unposed) images in, 3D Gaussians out, behind one `GSPipeline.from_pretrained` for every model. ```bash pip install --pre ffgs # inference pip install --pre "ffgs[render]" # + gsplat rasterisation (CUDA) ``` ```python from ffgs import GSPipeline, Views pipe = GSPipeline.from_pretrained("anysplat/default", device="cuda") out = pipe(Views(images)) # unposed images -> Gaussians + cameras out.gaussians.save_ply("scene.ply") # standard 3DGS ply ``` ## Models ffgs loads the authors' own checkpoints from their Hub repos (pinned to a commit and a sha256); it does not re-host weights. | model | weights | poses | license (weights) | |---|---|---|---| | [TokenGS](https://arxiv.org/abs/2604.15239) | [jiaweir/tokengs](https://huggingface.co/jiaweir/tokengs) | required | NVIDIA research license (non-commercial) | | [AnySplat](https://arxiv.org/abs/2505.23716) | [lhjiang/anysplat](https://huggingface.co/lhjiang/anysplat) | optional | non-commercial (fine-tuned from VGGT-1B, CC BY-NC 4.0) | Each model's page — sources, zoo entries, usage, and agreement with the official code: [docs/models](https://github.com/libgs/ffgs/blob/main/docs/models/index.md). ## Links - Code: [github.com/libgs/ffgs](https://github.com/libgs/ffgs) - Package: [pypi.org/project/ffgs](https://pypi.org/project/ffgs/) - Licenses: the ffgs core is Apache-2.0; vendored model code and all weights keep their own (several non-commercial) licenses — see the [README](https://github.com/libgs/ffgs#license).