Omega-Forge
An unofficial implementation for fine-tuning VGGT-Omega, a feed-forward 3D reconstruction model, on your own data, plus a standalone GPU-free 3D reconstruction viewer.
Omega-Forge is an unofficial training implementation for VGGT-Omega, written because the model's official training code was not yet available. Its training component takes a JSONL manifest of posed RGB-D frames and adapts VGGT-Omega's depth and camera heads to a new domain, with a held-out validation loop; the project is explicit that this is not a reproduction of the original authors' training procedure.
A separate, standalone visualization component imports neither PyTorch nor VGGT-Omega, so it can render any 3D reconstruction, from any source, on any machine with no GPU, unprojecting predicted depth into a point cloud with the camera path and view frustums built up frame by frame. Fine-tuned separately across six benchmarks with one unchanged recipe, depth and pose accuracy improve overall, and the project reports that simply not stretching a non-square source image during preprocessing improved the frozen model's depth accuracy more than fine-tuning did.
It is written in Python and MIT licensed.