arXiv preprint / cs.LG

DreOPD: Degraded-Reference Extrapolative On-Policy Distillation for Flow-matching Models

作者：Mingfeng Lin, Chengfei Cai, Lin Xu, Yuxiang Wei, Liang Han · 单位：Harbin Institute of Technology (Shenzhen), Zhejiang University, Harbin Institute of Technology · 会议/期刊：arXiv preprint · 方向：cs.LG · 发布日期：2026-08-11

2026-08-11生成模型强化学习
新标签打开 AlphaXiv新标签打开 Project
DreOPD: Degraded-Reference Extrapolative On-Policy Distillation for Flow-matching Models figure
AlphaXiv 中文概览（可滚动查看）