microsoft/MoGe

What it solves

MoGe addresses the challenge of recovering accurate 3D geometry from a single, open-domain image. It enables the estimation of metric point maps, metric depth maps, normal maps, and the camera's field of view (FOV) from a single forward pass, reducing the need for multiple specialized models.

How it works

MoGe uses a deep learning model (with versions ranging from ViT-S to ViT-G) to analyze a monocular image and predict its 3D structure. The latest version, MoGe-3, introduces self-guided sparse volumetric refinement to improve fine-grained geometry. The model can either estimate the camera FOV automatically or use a provided ground-truth FOV to increase accuracy. It supports flexible resolutions and aspect ratios, and can be optimized for speed using FP16 precision.

Who it’s for

This tool is designed for researchers and developers working in computer vision, 3D reconstruction, and augmented reality who need to extract precise metric 3D data from standard 2D photographs.

Highlights

• All-in-one estimation: Predicts point maps, depth maps, and normal maps simultaneously.
• Metric scale: Provides geometry with actual metric scale (available in MoGe-2 and MoGe-3).
• High efficiency: Achieves 60ms latency per image on high-end GPUs (A100/RTX3090).
• Panorama support: Includes an experimental extension for processing 360° equirectangular panorama images.
• Flexible output: Exports results as .glb or .ply files for use in 3D software.