MoGe-2

MoGe-2 is Microsoft's state-of-the-art model for monocular geometry estimation, released June 10, 2025 as the follow-up to MoGe (a CVPR 2025 Oral). From a single open-domain image it predicts metric-scale point maps, depth maps, normal maps, and camera FOV in one forward pass. Compared with MoGe-1 it adds metric scale prediction, comparable or better accuracy, visibly sharper geometry, high-quality normal map estimation, and lower inference latency (60ms per image on an A100 or RTX 3090 at ViT-L, FP16).

The code is MIT licensed (except the DINOv2 module, Apache 2.0 from Meta), with checkpoints published via the microsoft/MoGe GitHub repo. Use cases span 3D reconstruction, AR content creation, and robotics perception where real-world scale from a single camera frame matters.