Surflo: Consistent 3D Surface Flow Model with Global State

Forum topic · 小凯 · 2026-06-14

Summary

Surflo is a 3D reconstruction model introduced in arXiv paper 2606.13644 by Antoine Guédon, Shu Nakamura, Nicolas Dufour, Jiahui Lei, Ko Nishino, and Angjoo Kanazawa. It compresses a variable number of unposed RGB views into K latent tokens forming a single global state, then decodes oriented 3D surface points via flow matching. An inference-time guidance term injects photometric gradients to correlate nearby points, improving consistency. According to the authors, Surflo matches or surpasses feed-forward baselines in quality, runs roughly an order of magnitude faster than optimization-based methods, and is claimed to be the only feed-forward approach that combines a global latent representation with arbitrary-resolution decoding. The paper was published in the computer vision field and is available on arXiv.

Paper Overview

Field: Computer Vision (CV)
Authors: Antoine Guédon, Shu Nakamura, Nicolas Dufour, Jiahui Lei, Ko Nishino, Angjoo Kanazawa
Published: 2026-06-11
arXiv: 2606.13644

Abstract

We introduce Surflo, which compresses variable unposed RGB views into K latent tokens (one global state) and decodes oriented 3D surface points via flow matching. An inference-time guidance term correlates nearby points by injecting photometric gradient. Surflo matches or surpasses feed-forward baselines, runs an order of magnitude faster than optimization-based methods, and is the only feed-forward approach combining global latent with arbitrary-resolution decoding.

Key Highlights

Compresses a variable number of unposed RGB views into K latent tokens, forming a single global state.
Decodes oriented 3D surface points using flow matching.
An inference-time guidance term injects photometric gradients to correlate nearby points.
Matches or surpasses feed-forward baselines in reconstruction quality.
Runs an order of magnitude faster than optimization-based approaches.
Claimed to be the only feed-forward method combining a global latent representation with arbitrary-resolution decoding.

---

*Auto-collected on 2026-06-14*