FreeMatching Couples FLUX.2 and DINOv3 for Identity-Preserving Dense Correspondence
Summary
FreeMatching stitches FLUX.2-4B with DINOv3 and trains in two stages — supervised pre-training on 520K tracked video clips plus 360K synthetic Blender pairs, then teacher-guided refinement on image-editing-and-generation data without dense labels. On the new IEG-Bench reconstruction protocol it tops RoMa by 5.57 Full-MSE and UFM by 2.48, while staying competitive with UFM and RoMa on ETH3D, ScanNet-1500 and Sintel/KITTI.
Originally reported by huggingface.co
Read the original article →Original headline: FreeMatching Couples FLUX.2 and DINOv3 for Identity-Preserving Dense Correspondence