Beyond Spatio-Temporal Priors: A Generalizable Approach for Dense Correspondence Matching
FreeMatching targets identity matches where physical continuity no longer holds.
The paper argues that dense correspondence methods built on smooth motion or rigid geometry fail on image editing and reference-guided generation pairs. Its proposed FreeMatching model combines generative and semantic foundation representations, trained with supervision from classical datasets, tracked videos, and synthetic scenes. The authors report better correspondence on difficult IEG image pairs while staying competitive on classical benchmarks. They also present it as a metric for identity preservation, with scores said to correlate with human judgment. HF Daily Papers' note
The paper argues that dense correspondence methods built on smooth motion or rigid geometry fail on image editing and reference-guided generation pairs. Its proposed FreeMatching model combines generative and semantic foundation representations, trained with supervision from classical datasets, tracked videos, and synthetic scenes. The authors report better correspondence on difficult IEG image pairs while staying competitive on classical benchmarks. They also present it as a metric for identity preservation, with scores said to correlate with human judgment. HF Daily Papers' note
score 4