MoGe-3 by Microsoft Sets SOTA on 9 Benchmarks for High-Fidelity 3D Geometry from a Single Image
RexDouglass · x · 2026-08-01
Microsoft Research, in collaboration with Tsinghua University and USTC, released MoGe-3, a new monocular geometry estimation model. It introduces Self-Guided Sparse 3D Refinement (SSR), which lifts coarse point maps from a base model onto a sparse voxel shell and refines depth using lightweight sparse 3D convolutions.
By operating in 3D space rather than the 2D image plane, SSR avoids feature mixing across depth discontinuities, successfully recovering fine structures typically over-smoothed by 2D decoders. Across 9 zero-shot benchmarks, MoGe-3 achieves state-of-the-art global and local accuracy with major gains on strict fine-detail metrics, matching Depth Pro's boundary sharpness at a much lower resolution.
More from Multimodal
- MiniMax 3 passes Turing test for AI video: writes 'Hi' on chalkboard, now open source — Kyrannio · 2026-08-01
- Testing Doubao vs. Jimeng: $30 Gets You Only 3-6 AI Video Generations — oran_ge · 2026-08-01
- Experiment Shows Training AI Image Models at 512 Resolution + Upscaling Saves VRAM — More_Bid_2197 · 2026-08-01
- 8-Hour Debug of ComfyUI Black Images Uncovers PyTorch FP16 Overflow Bug — rtsitola · 2026-08-01
- Running Ideogram 4 Locally on Apple Silicon: Best Typography Workflow — DaLyon92x · 2026-08-01
- Breaking Image Boundaries: Creating 'Enterable' Interactive Video with Seedance 2.0 — Kyrannio · 2026-08-01