VibeToken: Scaling 1D Image Tokenizers and Autoregressive Models for Dynamic Resolution Generations

Researchers from Arizona State University and Huawei have unveiled VibeToken, a novel resolution-agnostic autoregressive (AR) image synthesis approach that challenges the dominance of diffusion models. At its core, VibeToken is a 1D Transformer-based image tokenizer that dynamically encodes images into a user-controllable sequence of 32-256 tokens, achieving a state-of-the-art efficiency and performance trade-off. This enables arbitrary resolution and aspect ratio generation without retraining, a major limitation of fixed-resolution AR models like LlamaGen.

Building on VibeToken, the team introduces VibeToken-Gen, a class-conditioned AR generator that synthesizes 1024x1024 images using only 64 tokens, achieving a 3.94 gFID score. In comparison, a diffusion-based state-of-the-art alternative requires 1,024 tokens and attains 5.87 gFID. VibeToken-Gen maintains constant 179G FLOPs regardless of resolution, a 63.4x efficiency improvement over LlamaGen's 11T FLOPs at 1024x1024. Accepted at CVPR'26, this work could accelerate production deployment of AR visual generative models.