FreeMatching: generalizable dense correspondence matching beyond spatio-temporal priors
hkuhk · hf · 2026-10-09
- FreeMatching tackles the limits of dense correspondence matching built on spatio-temporal priors (smooth motion, rigid geometry), which break down in image editing and reference-guided generation (IEG) where identity persists but physical continuity doesn't.
- The framework combines generative and semantic foundation representations with heterogeneous supervision from classical datasets, tracked videos, and synthetic scenes; teacher-guided iterative refinement improves IEG correspondence without dense annotations.
- A single model substantially improves correspondence on challenging IEG pairs, stays competitive on classical benchmarks, and works as a quantitative identity-preservation metric correlating with human judgment. Code is open-sourced.
More from Multimodal
- Iris 3B text-to-image samples: 1360x1024 renders in ~1 minute on Blackwell — b4silio · 2026-10-09
- GPT-Image x Fable collaboration yields stunning Iliad book 1-2 illustrations — aiamblichus · 2026-10-09
- The expensive part of AI video isn't generating it, it's regenerating it — Khalizo · 2026-10-09
- Synkinema: open-source MCP video editor lets AI agents and humans share one timeline — United_Inspector_653 · 2026-10-09
- Prompt Share: 'Digital Fracture' Template for Glitch-Art AI Images — azed_ai · 2026-10-09
- Creator releases 72 Krea 2 Asian character LoRAs trained on gravure idol photobooks — Fideltampan12345 · 2026-10-09