New 3D foundation model paper uses Riemannian flow matching to stay on the manifold
kwangmoo_yi · x · 2026-07-24
The paper Latent Riemannian Flow Matching for Geometry-Grounded 3D Foundation Models proposes a way to denoise in the latent space of VGGT so its 3D priors can be used more effectively.
The key technical move is to use Riemannian flow matching to keep the process on the manifold. The accompanying visual comparisons show context images, ground truth, and outputs against Gen3R and DepthSplat across multiple scenes, illustrating the geometry-grounded reconstruction quality.
Related event: Riemannian Flow Matching Powers New 3D Foundation Model(2 posts)→
More from Embodied
- BeingBeyond’s Being-M0.7 learns humanoid motion from video and beats prior methods on Unitree G1 — jiqizhixin · 2026-07-24
- SN44 launches a card-grading challenge that scores five visual quality signals — bittingthembits · 2026-07-24
- AMD’s next Gorgon Halo platform will raise unified memory to 192GB — ryanshrout · 2026-07-24
- AMD’s Ryzen AI Halo targets local AI apps with 128GB unified memory — ryanshrout · 2026-07-24
- Black Forest Labs launches Flux 3 with native audio video generation up to 20 seconds — The Decoder · 2026-07-24
- Robocurve aims to benchmark robots on real-world physical tasks — garrytan · 2026-07-24