Depth Anything V3

Depth Anything V3 is a transformer-based depth estimation model released November 14, 2025 with paper, project page, code, and weights all published. It moves beyond disparity-based depth to a unified depth-ray representation: one model handles monocular depth estimation, multi-view depth, pose-conditioned depth, camera pose estimation, and direct 3D Gaussian prediction for novel view synthesis. It ships in four sizes (Giant, Large, Base, Small) plus Metric and Monocular series.

Reported results claim state-of-the-art monocular depth with 44.3% better camera pose accuracy than VGGT, though independent reproduction is pending. Primary uses are robotics navigation, AR/VR depth mapping, computational photography, and scene understanding, with ComfyUI integration for depth-based creative workflows. Weights and training code are open.