Research map / Probabilistic and scientific modeling / Variational and generative sampling
Reweighted Flow Matching via Unbalanced Optimal Transport for Label-free Long-tailed Generation
Hyunsoo Song, Minjung Gim, Jaewoong Choi
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UOT-RFM, a label-free reweighted flow matching method using unbalanced optimal transport to counter majority bias in long-tailed generative modeling.
Problem
Standard flow matching over-represents dominant modes on long-tailed distributions, hurting tail-class fidelity and distributional accuracy.
Approach
Reweights flow matching training using unbalanced optimal transport so that tail-class samples are emphasized without requiring class labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add UOT-based inverse majority-score reweighting to flow matching | flow matching · CIFAR-10-LT (I=0.01, LT->LT) | FID: 14.25 -> 11.03 | Improved |
| add UOT-RFM on CIFAR-100-LT with I=0.001 | flow matching · CIFAR-100-LT (I=0.001, LT->LT) | FID: 31.83 -> 18.40 | Improved |
| use UOT-RFM instead of OT-CFM | flow matching · CIFAR-10-LT (I=0.01) class distribution | mean NCRE: 1.02 -> 0.40 | Improved |
| use UOT-RFM instead of UOT-CFM | flow matching · CIFAR-10-LT (I=0.01, LT->LT) | Precision: 0.73 -> 0.62 | Hurt |
| use UOT-RFM instead of UOT-CFM | flow matching · CIFAR-10-LT (I=0.01, LT->LT) | Recall: 0.29 -> 0.41 | Improved |
| increase correction order k from 1 to 8 (tau=2.0) (ablation) | UOT-RFM · CIFAR-10-LT (I=0.01, LT->LT) | FID: 13.77 -> 11.37 | Improved |
| increase tau from 2.0 to 6.0 (k=8) (ablation) | UOT-RFM · CIFAR-10-LT (I=0.01, LT->LT) | FID: 11.37 -> 12.41 | Hurt |
| increase correction order k (tau=2.0) in LT->Balanced (ablation) | UOT-RFM · CIFAR-10-LT (I=0.01, LT->Balanced) | FID: 25.02 -> 24.54 (k=4), 24.76 (k=8) | Mixed |
| use UOT-RFM (k=1) instead of OT-CFM on balanced CIFAR-100 | flow matching · CIFAR-100 (balanced) | FID: 6.14 -> 6.54 | Hurt |
| use UOT-RFM (k=1) instead of UOT-CFM on balanced CIFAR-10 | flow matching · CIFAR-10 (balanced) | FID: 3.62 -> 3.58 | No clear effect |
Experimental setup
Builds on: flow matching
Closest papers by meaning
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · similarity 0.89)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.88)
- Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling (ICLR 2025 · similarity 0.88)
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (NeurIPS 2025 · similarity 0.88)
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.88)
- Improving Flow Matching by Aligning Flow Divergence (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
