FreeMatching: generalizable dense correspondence beyond spatio-temporal priors
arXiv:2610.12421 (submitted 8 Oct 2026, accepted at NeurIPS 2026) introduces FreeMatching, a framework that combines generative and semantic foundation representations with heterogeneous supervision from classical datasets, tracked videos and synthetic scenes for dense correspondence matching in image editing and reference-guided generation (IEG), where smooth-motion and rigid-geometry priors break down. The authors report that a single model substantially improves correspondence quality on challenging IEG image pairs while staying competitive on classical benchmarks, with teacher-guided iterative refinement improving results without dense correspondence annotations. They also demonstrate its use as a quantitative metric for identity preservation, with scores said to correlate with human judgment; code is available.
It extends dense correspondence beyond classical geometric assumptions to image editing and reference-guided generation, and offers a way to score identity preservation that tracks human judgment.