Repository navigation
First release of VGGT-Splat: a controlled comparison of COLMAP → 3DGS and feed-forward VGGT → 3DGS (with and without a VGGT depth prior) on 6 scenes × 8 / 16 / 32 / all training frames. Findings, conclusions and reproduction steps are in the README; videos and figures are on the project page.
Trained models
All-frames models of every scene, 3DGS trained 7,000 steps with pose refinement (<scene>_<colmap|vggt>_all), in two formats:
.splat: base color only, for web viewers that read the antimatter15.splatformat..compressed.ply: PlayCanvas compressed PLY with spherical harmonics, e.g. for SuperSplat.
| Scene | COLMAP test PSNR | VGGT test PSNR |
|---|---|---|
| desk (self-captured) | 32.46 | 31.79 |
| lawn (self-captured) | 21.44 | 19.35 |
| room (self-captured) | 23.40 | 21.20 |
| garden (Mip-NeRF 360) | 26.34 | 21.79 |
| bicycle (Mip-NeRF 360) | 23.40 | 21.30 |
| counter (Mip-NeRF 360) | 27.25 | 25.04 |
PSNR is after test-time pose refinement, on the test frames both methods registered (lawn 22/23, room 14/17). Each model is in its own normalized world frame, so the COLMAP and VGGT models of a scene are not aligned with each other.
Self-captured data
vggt-splat_data_<scene>.zip (desk, lawn, room) holds the extracted frames/ and the exact splits/ used in the study; unzip inside data/. Mip-NeRF 360 scenes are not redistributed; see the README for setting them up.
Verify downloads with shasum -a 256 -c SHA256SUMS (macOS) or sha256sum -c SHA256SUMS (Linux).