Uncertainty-Gated Adapter for Visual Autoregressive Generation

Abstract

Visual AutoRegressive modeling (VAR) has recently revitalized autoregressive visual generation through next-scale prediction. However, its scale-level parallel prediction treats all tokens uniformly, despite their drastically different uncertainty. This limitation allows unstable tokens to dominate feature propagation, ultimately constraining the upper bound of nextscale generation quality. To address this weakness, we introduce UGA, an Uncertainty-Gated Adapter that enables tokenwise, uncertainty-aware refinement for VAR. UGA estimates token uncertainty using a lightweight probing network and converts it into a continuous gate that modulates a bottleneck adapter inserted between Transformer blocks. This selective refinement strengthens unstable tokens while preserving reliable ones, improving structural stability without altering the backbone architecture. Extensive experiments demonstrate that UGA consistently enhances semantic coherence and generation fidelity with <2% additional parameters, offering a simple yet effective direction for stabilizing next-scale autoregressive generation.

Authors 2

1. Yanran Xu Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Aachen,Germany

2. Xun Jia Su

   Jiangxi Agricultural University

   Affiliation as printed

   Jiangxi Agricultural University,Nanchang,China