VGGT-Ω

VGGT-Ω (VGGT-Omega) is a family of one-billion-parameter models, presented as an oral paper at CVPR 2026 (arXiv 2605.15195).¹ Feed-forward means the model computes its answer in one pass through the network.² Given a set of frames, it returns camera poses and intrinsics (lens settings such as focal length), per-pixel depth with confidence, which can be unprojected into a point cloud, and camera tokens.¹ Variants are trained at 512, 416 and 256 pixel resolution. One checkpoint adds a text-alignment embedding.¹

DeepReefMap, from EPFL and MIT, builds a semantic 3D point cloud, an ortho-mosaic and benthic cover statistics for a reef from handheld underwater video.² An ortho-mosaic is a stitched overhead image with perspective removed. Benthic cover is what covers the seabed, such as coral or sand.² VGGT-Ω is one of DeepReefMap's four reconstruction backends, and its README rates it high quality.² Unlike the streaming alternative, it looks at the whole clip at once, so GPU memory caps the length of video it can handle.² The README quotes upstream figures of roughly 6 GB for the model plus about 75 MB per frame on an A100.^1,2

The code is on GitHub.¹ The weights are at facebook/VGGT-Omega on Hugging Face, behind an access request that is reviewed automatically.^1,2 DeepReefMap names the licence as the FAIR Noncommercial Research License and keeps the backend as an opt-in extra for that reason.² The VGGT-Ω repository points readers to its LICENSE file and lists updates from September 2026.¹ The model runs on NVIDIA, AMD and Apple Silicon GPUs.²