Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping

TL;DR

This paper introduces Super-Linear Advantage Shaping (SLAS), a novel reinforcement learning method for text-to-image models that enhances training efficiency, robustness, and performance by reshaping policy updates based on advantage-dependent geometry.

Contribution

SLAS extends the Fisher-Rao metric with advantage-dependent weighting, improving policy updates and mitigating reward hacking in post-training T2I models.

Findings

SLAS outperforms DanceGRPO across multiple benchmarks.

Faster training dynamics and better out-of-domain performance.

Enhanced robustness to model scaling and reduced reward hacking.

Abstract

Recently, post-training methods based on reinforcement learning, with a particular focus on Group Relative Policy Optimization (GRPO), have emerged as the robust paradigm for further advancement of text-to-image (T2I) models. However, these methods are often prone to reward hacking, wherein models exploit biases in imperfect reward functions rather than yielding genuine performance gains. In this work, we identify that normalization could lead to miscalibration and directly removing the prompt-level standard deviation term yields an optimal policy ascent direction that is linear in the advantage but still limits the separation of genuine signals from noise. To mitigate the above issues, we propose Super-Linear Advantage Shaping (SLAS) by revisiting the functional update from an information geometry perspective. By extending the Fisher-Rao information metric with advantage-dependent weighting, SLAS introduces a non-linear geometric structure that reshapes the local policy space. This design relaxes constraints along high-advantage directions to amplify informative updates, while tightening those in low-advantage regions to suppress illusory gradients. In addition, batch-level normalization is applied to stabilize training under varying reward scales. Extensive evaluations demonstrate that SLAS consistently surpasses the DanceGRPO baseline across multiple backbones and benchmarks. In particular, it yields faster training dynamics, improved out-of-domain performance on GenEval and UniGenBench++, and enhanced robustness to model scaling, while mitigating reward hacking and preserving semantic and compositional fidelity in generations.