Qwen logo TEAI logo

UniAR: Unified Multimodal Autoregressive Modeling with Shared Context

Visual Tokenizer is Key to Unification

ICML 2026


Wujian Peng1,2,*,#    Lingchen Meng3,*,‡    Yuxuan Cai3    Xianwei Zhuang3    Yuhuan Yang3    Rongyao Fang3    Chenfei Wu3    Junyang Lin3    Zuxuan Wu1,2,†    Shuai Bai3,†
* Equal contributions     # Work done during internship at Qwen     † Corresponding authors     ‡ Project lead
1 Institute of Trustworthy Embodied AI, Fudan University     2 Shanghai Innovation Institute
3 Qwen Team, Alibaba Group
Arxiv GitHub Checkpoints

UniAR teaser

UniAR is a unified autoregressive multimodal model that handles image understanding, image generation, and image editing in a single Transformer. Unlike prior unified models that rely on two separate visual tokenizers (splitting the representation space), UniAR uses a single discrete visual tokenizer as the key bridge between understanding and generation, enabling a shared context in which the model can directly interpret its own generated visual tokens without additional re-encoding.

Key design choices:

  • Multi-level BSQ tokenizer — fuses shallow (low-level detail) and deep (high-level semantic) visual features via lookup-free Binary Spherical Quantization, scaling the effective vocabulary to 264 codes with minimal overhead.
  • Parallel bitwise prediction — jointly predicts spatially grouped, multi-level visual codes per AR step, achieving a 32x visual compression ratio (a 1024×1024 image needs only 256 AR tokens).
  • DiT-based visual decoder — an SD3-medium transformer with semantic visual feature injection that reconstructs high-fidelity images from discrete visual tokens, with resolution upsampling support.

Highlights

  • True unification via shared context. A single visual tokenizer is used for both understanding and generation, allowing UniAR to interpret its own generated visual tokens directly.
  • Bitwise visual tokenization at scale. Lookup-free BSQ quantization expands the effective visual vocabulary with low overhead while preserving semantic alignment.
  • Multi-level visual features. Hierarchical feature fusion retains both high-level semantics and fine-grained details, which is especially important for text rendering and editing.
  • Fast autoregressive generation. Parallel bitwise prediction and a diffusion-based visual decoder with resolution upsampling reduce sequence length and accelerate image synthesis.
  • Strong multimodal performance. UniAR achieves state-of-the-art or highly competitive results on image generation, image editing, OCR-heavy understanding, and long-text rendering.

Method Overview

UniAR architecture
UniAR consists of three core components: a unified visual tokenizer that discretizes semantic visual features into shared bitwise tokens, a unified autoregressive backbone that jointly models text and visual tokens, and a diffusion-based visual decoder that reconstructs high-fidelity images from predicted visual tokens.

Interleaved Generation and Understanding

UniAR interleaved generation-understanding example
UniAR unifies generation and understanding in the same discrete visual space. This enables an emergent interleaved capability: after generating an image, the model can answer follow-up questions about its own output without re-encoding the image.

BibTeX

@article{peng2026unified,
  title={Unified Multimodal Autoregressive Modeling with Shared Context-Visual Tokenizer is Key to Unification},
  author={Peng, Wujian and Meng, Lingchen and Cai, Yuxuan and Zhuang, Xianwei and Yang, Yuhuan and Fang, Rongyao and Wu, Chenfei and Lin, Junyang and Wu, Zuxuan and Bai, Shuai},
  journal={arXiv preprint arXiv:2606.18249},
  year={2026}
}