LISA

LISA: Locality-Informed Speculative Decoding for Accelerating Autoregressive Image Generation

ECCV 2026

Ying Li^1*, Siyong Jian^1*, Zhaode Wang², Zhiwen Chen², Chengfei Lv², Huan Wang^1†

¹ Westlake University   ·   ² Alibaba Group

^* Equal contribution   ·   ^† Corresponding author: wanghuan [at] westlake [dot] edu [dot] cn

Code Paper Apache 2.0

LISA accelerates visual autoregressive generation with locality-informed speculative decoding, achieving up to 4.12× speedup on Lumina-mGPT-7B, 2.11× on LlamaGen-XL-T2I, and 1.53× on LlamaGen-XL-C2I, with competitive generation quality.

Abstract

Autoregressive (AR) models deliver high-quality image generation but suffer from severe inference latency due to sequential decoding. While speculative decoding (SD) successfully accelerates large language models, we demonstrate that state-of-the-art methods like EAGLE-3 provide limited gains for visual AR models. We attribute this to the intrinsic nature of visual tokens: unlike text, their probability distributions are flat and non-discriminative, which hinders drafter–target alignment and causes verification to collapse under sampling. To bridge this gap, we propose LISA, a Locality-Informed Speculative framework for Autoregressive image generation. LISA overcomes alignment barriers via two key components: (i) Locality-Informed Distillation, which aligns the drafter by prioritizing supervision on structured uncertainty; and (ii) Geometry-Aware Soft Verification, which leverages embedding proximity and target confidence to enable more permissive yet safe token acceptance. Experiments across T2I and C2I tasks demonstrate that LISA improves the speed–quality trade-off, achieving up to 4.12× acceleration and outperforming recent competitive methods.

BibTeX

@inproceedings{li2026lisa,
  title={LISA: Locality-Informed Speculative Decoding for Accelerating Autoregressive Image Generation},
  author={Li, Ying and Jian, Siyong and Wang, Zhaode and Chen, Zhiwen and Lv, Chengfei and Wang, Huan},
  booktitle={ECCV},
  year={2026}
}

More content (method figure, results, visualizations) coming soon.