New RL Framework Boosts Few-Step Flow-Map Image Generators

Current few-step flow-map generators, like consistency models, are highly efficient for tasks such as text-to-image generation, as they directly learn long-range transport maps. However, their deterministic nature makes them challenging to optimize using reinforcement learning (RL) post-training methods, which typically require stochastic trajectories and clear likelihood ratios. Existing stochasticization techniques are not directly applicable to these long-range flow maps. A new framework, Flow-Map GRPO, has been introduced to address this limitation. It provides an an online RL post-training mechanism specifically designed for deterministic few-step flow-map generators. The core innovation is Anchored Stochastic Flow Map Composition (ASFMC), which injects randomness through anchor-based conditional resampling while preserving the original deterministic flow map's marginal probability path. Experiments with FLUX-based text-to-image generators, including MeanFlow and sCM, demonstrated that Flow-Map GRPO significantly enhances pretrained deterministic models. The improvements were observed across various metrics, including reward-based, perceptual, and task-level evaluations, proving that RL can effectively align these models without requiring architectural changes or retraining them as native stochastic models.