ReFlowSET: Representation-Aligned Latent Flow Matching for SAR-to-EO Image Translation
ReFlowSET treats the SAR-to-EO latent codec as a core design choice, not a borrowed default.
The paper proposes a conditional latent flow-matching system that first audits SAR and EO reconstruction quality to choose its codec. It then trains a smaller conditional DiT from scratch, using dual-stream SAR conditioning and feature refinement. During training, noisy EO features are aligned to clean EO representations from a frozen vision foundation model, with no added inference cost. The authors report state-of-the-art results on QXS-SAROPT and SAR2Opt and say code and pretrained weights are public. HF Daily Papers' note
The paper proposes a conditional latent flow-matching system that first audits SAR and EO reconstruction quality to choose its codec. It then trains a smaller conditional DiT from scratch, using dual-stream SAR conditioning and feature refinement. During training, noisy EO features are aligned to clean EO representations from a frozen vision foundation model, with no added inference cost. The authors report state-of-the-art results on QXS-SAROPT and SAR2Opt and say code and pretrained weights are public. HF Daily Papers' note
score 3