Flow-GRPO: Training Flow Matching Models via Online RL
Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, Wanli Ouyang
2025-05-09
Summary
This paper talks about Flow-GRPO, a new method that makes AI models better at creating images from text by combining two advanced techniques: flow matching and online reinforcement learning.
What's the problem?
The problem is that generating high-quality images from text descriptions is really hard for AI, especially if you want the process to be both fast and accurate. Existing methods can be slow or not always produce the best results, which limits their usefulness.
What's the solution?
The researchers developed Flow-GRPO, which links flow matching models with online reinforcement learning. They use a special mathematical trick to convert one type of equation to another and reduce noise in the process, making it easier and more efficient for the AI to generate images that match the text.
Why it matters?
This matters because it means AI can create better and faster images from text, which is useful for art, design, entertainment, and even education. Improving this technology makes creative tools more powerful and accessible for everyone.
Abstract
Flow-GRPO combines online reinforcement learning with flow matching models through an ODE-to-SDE conversion and denoising reduction, improving sampling efficiency and performance across text-to-image tasks.