ZipAR: Parallel Auto-regressive Image Generation through Spatial Locality

TL;DR

ZipAR introduces a parallel decoding framework for auto-regressive image generation that leverages local spatial structures to significantly reduce the number of forward passes, boosting efficiency without retraining.

Contribution

It presents a training-free, plug-and-play method that enables parallel token decoding in AR image generation by exploiting local spatial dependencies.

Findings

Reduces forward passes by up to 91% on Emu3-Gen.

No retraining required for the acceleration.

Improves generation efficiency significantly.

Abstract

In this paper, we propose ZipAR, a training-free, plug-and-play parallel decoding framework for accelerating auto-regressive (AR) visual generation. The motivation stems from the observation that images exhibit local structures, and spatially distant regions tend to have minimal interdependence. Given a partially decoded set of visual tokens, in addition to the original next-token prediction scheme in the row dimension, the tokens corresponding to spatially adjacent regions in the column dimension can be decoded in parallel, enabling the ``next-set prediction'' paradigm. By decoding multiple tokens simultaneously in a single forward pass, the number of forward passes required to generate an image is significantly reduced, resulting in a substantial improvement in generation efficiency. Experiments demonstrate that ZipAR can reduce the number of model forward passes by up to 91% on the Emu3-Gen model without requiring any additional retraining. Code is available here: https://github.com/ThisisBillhe/ZipAR.