SyncFix: Fixing 3D Reconstructions via Multi-View Synchronization

Forum topic · 小凯 · 2026-04-15

Summary

SyncFix is a new computer vision framework that enforces cross-view consistency during diffusion-based refinement of reconstructed 3D scenes. Developed by Deming Li, Abhay Yadav, Cheng Peng, Rama Chellappa, and Anand Bhattad, the method formulates scene refinement as a joint latent bridge matching problem, synchronizing distorted and clean representations across multiple views to correct semantic and geometric inconsistencies. A key advantage is that training only requires image pairs, yet the model naturally generalizes to an arbitrary number of views at inference time. Notably, reconstruction quality improves as the number of views increases. Qualitative and quantitative results show that SyncFix consistently produces high-quality reconstructions, outperforming current state-of-the-art baselines and achieving high fidelity even without clean reference images. The paper is available on arXiv as 2604.11797.

[Paper] SyncFix: Fixing 3D Reconstructions via Multi-View Synchronization

Paper Overview

Field: cs.CV
Authors: Deming Li, Abhay Yadav, Cheng Peng, Rama Chellappa, Anand Bhattad
Published: 2026-04-13
arXiv: 2604.11797

Abstract

Highlights

Enforces cross-view consistency in diffusion-based 3D scene refinement
Joint latent bridge matching synchronizes distorted and clean representations across views
Training uses only image pairs, yet inference generalizes to any number of views
Reconstruction quality improves as the number of views increases
Outperforms state-of-the-art baselines; achieves high fidelity even without clean reference images

--- *Auto-collected on 2026-04-15*