Dense correspondence matching has long relied on simplifying assumptions about smooth motion and rigid geometry. According to a new research paper, these spatio-temporal priors work well for classical tasks but fail in image editing and reference-guided generation (IEG), where transformations can preserve visual identity while breaking physical continuity.
To address this gap, researchers introduce FreeMatching, a generalizable framework that combines generative and semantic foundation representations with heterogeneous supervision drawn from classical datasets, tracked videos, and synthetic scenes. The approach also employs teacher-guided iterative refinement, which improves correspondence in IEG settings without requiring dense correspondence annotations.
Experiments show that a single FreeMatching model substantially improves correspondence quality on challenging IEG image pairs while maintaining competitive performance on classical benchmarks. The authors further demonstrate its utility as a quantitative metric for evaluating identity preservation, with scores that correlate with human judgment.
The code is available at https://github.com/luping-liu/FreeMatching.