FreqExit

FreqExit: Enabling Early-Exit Inference for Visual Autoregressive Models via Frequency-Aware Guidance

NeurIPS 2025

Ying Li¹, Chengfei Lv², Huan Wang¹

¹ Westlake University   ·   ² Alibaba Group

*Corresponding author: wanghuan [at] westlake [dot] edu [dot] cn

Code Apache 2.0

FreqExit bridges step-wise generation and early-exit acceleration, achieving up to 2× speedup with negligible quality degradation.

Abstract

FreqExit is a dynamic inference framework for Visual AutoRegressive (VAR) models, which decode from coarse structures to fine details. Existing methods fail on VAR due to the absence of semantic stability and smooth representation transitions. FreqExit addresses this by recognizing that high-frequency details essential to visual quality tend to emerge in later decoding stages. On ImageNet 256×256, FreqExit achieves up to 2× speedup with only minor degradation, and delivers 1.3× acceleration without perceptible quality loss. This enables runtime-adaptive acceleration within a unified model, offering a favorable trade-off between efficiency and fidelity for practical and flexible deployment.

Overview of our FreqExit method

Main Results

More Visualizations

Acknowledgement

This work builds upon the foundations of prior open-source efforts, including VAR, CoDe, and LayerSkip. We sincerely thank the authors for their excellent contributions to the research community.

BibTeX

@inproceedings{li2025freqexit,
  title={FreqExit: Enabling Early-Exit Inference for Visual Autoregressive Models via Frequency-Aware Guidance},
  author={Li, Ying and Lv, Chengfei and Wang, Huan},
  booktitle={NeurIPS},
  year={2025}
}