Wujian Peng, Lingchen Meng, Yuxuan Cai, Xianwei Zhuang, Yuhuan Yang, Rongyao Fang, Chenfei Wu, Junyang Lin et al.
UniAR is a unified autoregressive model that integrates image understanding and generation by sharing a single discrete visual tokenizer, overcoming the limitations of separate tokenizers and achieving state-of-the-art performance.
Existing unified multimodal models use different visual tokenizers for understanding and generation, splitting the representation space and hindering true unification. They also suffer from inefficiency due to re-encoding between tokenizers during generation.
UniAR adapts a pretrained vision encoder with multi-level feature fusion and lookup-free bitwise quantization to build a single discrete visual tokenizer. This tokenizer preserves both high-level semantics and low-level details while scaling the visual vocabulary at minimal cost. The unified autoregressive model uses parallel bitwise prediction to jointly predict spatially grouped multi-level visual codes, reducing sequence length and accelerating generation. A diffusion-based visual decoder then reconstructs high-fidelity images from the discrete tokens.
Through large-scale pretraining, supervised fine-tuning, and reinforcement learning, UniAR achieves state-of-the-art performance on image generation and editing, while remaining competitive on multimodal understanding benchmarks.