Depth Anything 3 predicts spatially consistent geometry from any number of images, with or without known poses. It argues for minimal modelling: a plain transformer (a vanilla DINO encoder) as backbone, and a single depth-plus-ray prediction target in place of multi-task heads. Training uses a teacher-student scheme, and all models are trained only on public academic datasets.
Findings
- On the authors’ new visual geometry benchmark it beats VGGT by an average of 44.3% in camera pose accuracy and 25.1% in geometric accuracy (abstract).
- It also beats Depth Anything 2 on monocular depth.
Limits
- Abstract only so far. The benchmark is the authors’ own.
Relevance
The predecessor, Depth Anything V2, is the pseudo-depth source inside several underwater splatting papers (RUSplatting) and the model cai-2025-underwater-metric-depth fine-tunes. If DA3 replaces it, the same underwater domain gap needs testing again. See foundation models for robotics and wang-2025-vggt.
Links
- arXiv · project page
- Subfield: Perception