SDE-Consistent Sampling for RL Post-Training of Flow-Matching Models

A recent study published on arXiv introduces "Precise," a groundbreaking method designed to enhance flow-matching models via post-training reinforcement learning. The approach substitutes the traditional deterministic reverse-time Ordinary Differential Equation with a Stochastic Differential Equation, enabling the development of a stochastic policy for online reinforcement learning. Researchers focus on two key aspects: optimizing stochastic exploration and efficiently discretizing the SDE with fewer steps. They assess the balance between exploration and stability, resulting in a new SDE schedule that significantly improves prompt alignment and perceptual quality in diffusion and flow-matching generative models.