| Contents | Flow Matching (FM) is a framework that learns a continuous flow defined by an ODE vector field to transport a source distribution to a target data distribution. In the Conditional Flow Matching (CFM) view source-target pairs are constructed and the vector field is regressed along interpolation paths. Recent studies have explored better source–target couplings to straighten trajectories and reduce numerical stiffness with OT-CFM based on mini-batch Optimal Transport (OT) as a representative example. However, OT-CFM still suffers from majority bias in long-tailed distributions where the training signal is concentrated on majority classes. To overcome this issue we propose Unbalanced Optimal Transport–Reweighted Flow Matching (UOT-RFM). The main idea is to estimate a label-free majority score from the target marginal of source-fixed UOT and use it as an inverse weight in the loss thereby balancing the training signal toward minority classes. The weighting exponent k controls the strength of correction: when k=1 UOT-RFM exactly recovers the target distribution and when k>1 it further emphasizes minority classes. Our method can be easily integrated into existing CFM/OT-CFM pipelines without additional labels or complex resampling. Experiments on CIFAR 10/100-LT and other long-tailed benchmarks show that UOT-RFM improves recall F1 score and class proportion alignment for minority classes while maintaining competitive FID compared to baselines. In this talk, I will first review FM and OT-CFM then explain the design and theoretical guarantee of UOT-RFM followed by experimental results. This work is a joint collaboration with Dr. Hyunsoo Song (NIMS) and Prof. Jaewoong Choi (SKKU) |
