VibeToken: Scaling 1D Image Tokenizers and Autoregressive Models for Dynamic Resolution Generations
Abstract
We introduce an efficient, resolution-agnostic autoregres002sive (AR) image synthesis approach that generalizes to ar003
bitrary resolutions and aspect ratios, narrowing the gap
004 to diffusion models at scale. At its core is VibeToken, a
005 novel resolution-agnostic 1D Transformer-based image tok006
enizer that encodes images into a dynamic, user-controllable
007 sequence of 32–256 tokens, achieving a state-of-the-art effi008
ciency and performance trade-off. Building on VibeToken,
009 we present VibeToken-Gen, a class-conditioned AR gener010
ator with out-of-the-box support for arbitrary resolutions
while requiring significantly fewer compute resources. No- 011
tably, VibeToken-Gen synthesizes 1024×1024 images us- 012
ing only 64 tokens and achieves 3.94 gFID; by compari- 013
son, a diffusion-based state-of-the-art alternative requires 014
1,024 tokens and attains 5.87 gFID. In contrast to fixed- 015
resolution AR models such as LlamaGen—whose inference 016
FLOPs grow quadratically with resolution (≈11T FLOPs at 017
1024×1024)—VibeToken-Gen maintains a constant 179G 018
FLOPs (63.4× efficient) independent of resolution. We hope 019
VibeToken can help unlock the wide adoption of AR visual 020
generative models in production use cases.
Authors
- Maitreya Patel
- Jingtao Li
- Weiming Zhuang
- Yezhou Yang
- Lingjuan Lyu
Venue
CVPR
Date
2026