Spaces:
Running on Zero
Apply for a GPU community grant: Academic project
ffgs is an open-source library (https://github.com/libgs/ffgs) that gives feed-forward 3D Gaussian Splatting models one standard inference interface. This Space is its interactive demo: upload a few photos or a short video, and a feed-forward model (AnySplat by default) reconstructs a 3D Gaussian scene in under a minute, shown in an in-browser viewer with progressive loading, an orbit video and PLY download.
It currently runs on ZeroGPU, which works, but one reconstruction (model forward + rendering the orbit video + building the level-of-detail set) needs 60-120 s of GPU time, so anonymous and free visitors run out of daily quota after one or two tries. A dedicated GPU (T4 or A10G, 24 GB preferred for larger view counts) would let anyone try the models without logging in.
The code, weights loading and demo are all public; model licenses are shown in the UI. It is research tooling for the 3D vision community, not a commercial service.