VAR RL Done Right: Tackling Asynchronous Policy Conflicts in Visual Autoregressive Generation

TL;DR

This paper introduces a novel reinforcement learning framework for visual autoregressive models that effectively manages asynchronous policy conflicts, leading to improved sample quality and alignment.

Contribution

It proposes an enhanced Group Relative Policy Optimization method with three components to address asynchronous conflicts in VAR models during RL training.

Findings

Significant improvements in sample quality over baseline

Enhanced objective alignment in VAR models

Robust and effective optimization demonstrated

Abstract

Visual generation is dominated by three paradigms: AutoRegressive (AR), diffusion, and Visual AutoRegressive (VAR) models. Unlike AR and diffusion, VARs operate on heterogeneous input structures across their generation steps, which creates severe asynchronous policy conflicts. This issue becomes particularly acute in reinforcement learning (RL) scenarios, leading to unstable training and suboptimal alignment. To resolve this, we propose a novel framework to enhance Group Relative Policy Optimization (GRPO) by explicitly managing these conflicts. Our method integrates three synergistic components: 1) a stabilizing intermediate reward to guide early-stage generation; 2) a dynamic time-step reweighting scheme for precise credit assignment; and 3) a novel mask propagation algorithm, derived from principles of Reward Feedback Learning (ReFL), designed to isolate optimization effects both spatially and temporally. Our approach demonstrates significant improvements in sample quality and objective alignment over the vanilla GRPO baseline, enabling robust and effective optimization for VAR models.