FreeMatching: A Generalizable Framework for Dense Correspondence in Image Editing and Reference-Guided Generation

Forum topic · 小凯 · 2026-10-11

Summary

FreeMatching is a generalizable dense correspondence matching framework introduced in an arXiv paper (2610.12421) by Luping Liu, Bingyi Kang, Yifan Wang, and Dong Xu. Dense correspondence matching has traditionally relied on spatio-temporal priors such as smooth motion and rigid geometry, which fail in image editing and reference-guided generation (IEG) where transformations preserve visual identity while breaking physical continuity. FreeMatching combines generative and semantic foundation representations with heterogeneous supervision drawn from classical datasets, tracked videos, and synthetic scenes. A teacher-guided iterative refinement further improves correspondence in IEG without requiring dense correspondence annotations. Experiments show that a single FreeMatching model substantially improves correspondence quality on challenging IEG image pairs while remaining competitive on classical benchmarks. The authors also demonstrate its use as a quantitative metric for identity preservation, with scores correlating with human judgments. Code is available on GitHub.

Paper Overview

Field: Computer Vision Authors: Luping Liu, Bingyi Kang, Yifan Wang, Dong Xu Published: 2026-10-08 arXiv: 2610.12421

Abstract

Dense correspondence matching has historically been bounded by simplifying spatio-temporal priors, such as smooth motion and rigid geometry. While effective for classical tasks, these assumptions break down in image editing and reference-guided generation (IEG), where transformations can preserve visual identity while breaking physical continuity. To establish identity-preserving correspondence across such transformations, the authors introduce FreeMatching, a generalizable framework combining generative and semantic foundation representations with heterogeneous supervision from classical datasets, tracked videos, and synthetic scenes.

Key contributions

Beyond spatio-temporal priors: addresses dense correspondence for transformations that preserve visual identity but violate smooth motion / rigid geometry assumptions.
Heterogeneous supervision: leverages classical datasets, tracked videos, and synthetic scenes together with generative and semantic foundation representations.
Teacher-guided iterative refinement: improves correspondence in IEG without dense correspondence annotations.
Identity-preservation metric: scores correlate with human judgments, enabling quantitative evaluation of identity preservation in generative models.

Results

A single FreeMatching model substantially improves correspondence quality on challenging IEG image pairs while remaining competitive on classical benchmarks.

Code: https://github.com/luping-liu/FreeMatching

*Auto-collected on 2026-10-11*