index.html

<!DOCTYPE html>
<html lang="en">

<head>
  <meta charset="utf-8">
  <meta name="description" content="UniAR: Unified Multimodal Autoregressive Modeling with Shared Context --- Visual Tokenizer is Key to Unification">
  <meta name="keywords" content="UniAR, unified multimodal model, autoregressive modeling, visual tokenizer, image generation, image editing, multimodal understanding">
  <meta name="viewport" content="width=device-width, initial-scale=1">
  <title>UniAR</title>
  <link rel="icon" href="data:,">

  <!-- Google tag (gtag.js) -->
  <script async src="https://www.googletagmanager.com/gtag/js?id=G-562412C65F"></script>
  <script>
    window.dataLayer = window.dataLayer || [];
    function gtag(){dataLayer.push(arguments);}
    gtag('js', new Date());
    gtag('config', 'G-562412C65F');
  </script>

  <link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro">
  <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bulma@0.9.1/css/bulma.min.css">
  <link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
  <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.1/css/all.min.css">
  <link rel="stylesheet" href="./static/css/index.css">
</head>

<body>

  <!-- ===== Logo Bar ===== -->
  <div class="logo-bar">
    <div class="container is-max-desktop">
      <img src="./figs/qwen-logo.png" alt="Qwen logo" class="logo-left">
      <img src="./figs/TEAI_logo.png" alt="TEAI logo" class="logo-right">
    </div>
  </div>

  <!-- ===== Hero ===== -->
  <section class="hero">
    <div class="hero-body">
      <div class="container is-max-desktop">
        <div class="columns is-centered">
          <div class="column has-text-centered">

            <h1 class="title is-2 publication-title">
              UniAR: Unified Multimodal Autoregressive Modeling with Shared Context
            </h1>
            <p class="publication-subtitle">Visual Tokenizer is Key to Unification</p>
            <p class="publication-venue">ICML 2026</p>

            <br>

            <div class="is-size-5 publication-authors authors-block">
              <span class="author-block"><a href="https://wjpoom.github.io/" target="_blank">Wujian Peng</a><sup>1,2,*,#</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block"><a href="https://scholar.google.com/citations?user=pUl9H8gAAAAJ&hl=en" target="_blank">Lingchen Meng</a><sup>3,*,&#8225;</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block">Yuxuan Cai<sup>3</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block">Xianwei Zhuang<sup>3</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block">Yuhuan Yang<sup>3</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block">Rongyao Fang<sup>3</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block">Chenfei Wu<sup>3</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block">Junyang Lin<sup>3</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block"><a href="https://zxwu.azurewebsites.net/" target="_blank">Zuxuan Wu</a><sup>1,2,&dagger;</sup>&nbsp;&nbsp;&nbsp;</span>
              <span class="author-block"><a href="https://scholar.google.com/citations?user=ylhI1JsAAAAJ&hl=zh-CN" target="_blank">Shuai Bai</a><sup>3,&dagger;</sup></span>
            </div>

            <div class="is-size-6 publication-authors corresponding-note">
              <span><sup>*</sup> Equal contributions &nbsp;&nbsp;&nbsp; <sup>#</sup> Work done during internship at Qwen &nbsp;&nbsp;&nbsp; <sup>&dagger;</sup> Corresponding authors &nbsp;&nbsp;&nbsp; <sup>&#8225;</sup> Project lead</span>
            </div>

            <div class="is-size-6 publication-authors publication-affiliations">
              <span class="author-block"><sup>1</sup> Institute of Trustworthy Embodied AI, Fudan University</span>&nbsp;&nbsp;&nbsp;&nbsp;
              <span class="author-block"><sup>2</sup> Shanghai Innovation Institute</span>
            </div>
            <div class="is-size-6 publication-authors publication-affiliations">
              <span class="author-block"><sup>3</sup> Qwen Team, Alibaba Group</span>
            </div>

            <div class="publication-links">
              <span class="link-block">
                <a href="https://arxiv.org/pdf/2606.18249" target="_blank">
                  <span class="icon"><i class="ai ai-arxiv"></i></span>
                  <span>Arxiv</span>
                </a>
              </span>
              <span class="link-block">
                <a href="https://github.com/ShareLab-SII/UniAR" target="_blank">
                  <span class="icon"><i class="fab fa-github"></i></span>
                  <span>GitHub</span>
                </a>
              </span>
              <span class="link-block">
                <a href="https://huggingface.co/collections/ShareLab-SII/uniar" target="_blank">
                  <span class="icon"><i class="fas fa-cube"></i></span>
                  <span>Checkpoints</span>
                </a>
              </span>
            </div>

            <br>

          </div>
        </div>
      </div>
    </div>

    <!-- Teaser + Abstract -->
    <div class="container is-max-desktop abstract-block">
      <div class="has-text-centered">
        <img id="teaser" width="88%" src="./figs/teaser.jpg" alt="UniAR teaser">
        <div class="content has-text-justified abstract-text">
          <p>
            <strong>UniAR</strong> 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 <strong>single discrete visual tokenizer</strong> 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.
          </p>
          <p><strong>Key design choices:</strong></p>
          <ul>
            <li><strong>Multi-level BSQ tokenizer</strong> — fuses shallow (low-level detail) and deep (high-level semantic) visual features via lookup-free Binary Spherical Quantization, scaling the effective vocabulary to 2<sup>64</sup> codes with minimal overhead.</li>
            <li><strong>Parallel bitwise prediction</strong> — 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).</li>
            <li><strong>DiT-based visual decoder</strong> — an SD3-medium transformer with semantic visual feature injection that reconstructs high-fidelity images from discrete visual tokens, with resolution upsampling support.</li>
          </ul>
        </div>
      </div>
    </div>
  </section>

  <!-- ===== Highlights ===== -->
  <section class="section section-compact" id="highlights">
    <div class="container is-max-desktop">
      <h2 class="title is-3 has-text-centered">Highlights</h2>
      <span class="section-divider"></span>
      <div class="highlight-box content">
        <ul>
          <li><strong>True unification via shared context.</strong> A single visual tokenizer is used for both understanding and generation, allowing UniAR to interpret its own generated visual tokens directly.</li>
          <li><strong>Bitwise visual tokenization at scale.</strong> Lookup-free BSQ quantization expands the effective visual vocabulary with low overhead while preserving semantic alignment.</li>
          <li><strong>Multi-level visual features.</strong> Hierarchical feature fusion retains both high-level semantics and fine-grained details, which is especially important for text rendering and editing.</li>
          <li><strong>Fast autoregressive generation.</strong> Parallel bitwise prediction and a diffusion-based visual decoder with resolution upsampling reduce sequence length and accelerate image synthesis.</li>
          <li><strong>Strong multimodal performance.</strong> UniAR achieves state-of-the-art or highly competitive results on image generation, image editing, OCR-heavy understanding, and long-text rendering.</li>
        </ul>
      </div>
    </div>
  </section>

  <!-- ===== Method Overview ===== -->
  <section class="section section-compact" id="method">
    <div class="container is-max-desktop has-text-centered">
      <h2 class="title is-3">Method Overview</h2>
      <span class="section-divider"></span>
      <img class="figure-image" src="./figs/arch.png" alt="UniAR architecture">
      <div class="figure-caption">
        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.
      </div>
    </div>
  </section>

  <!-- ===== Interleaved Generation & Understanding ===== -->
  <section class="section section-compact" id="shared-context">
    <div class="container is-max-desktop has-text-centered">
      <h2 class="title is-3">Interleaved Generation and Understanding</h2>
      <span class="section-divider"></span>
      <img class="figure-image" src="./figs/decoder.png" alt="UniAR interleaved generation-understanding example">
      <div class="figure-caption">
        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.
      </div>
    </div>
  </section>

  <!-- ===== BibTeX ===== -->
  <section class="section bibtex-section" id="BibTeX">
    <div class="container is-max-desktop content">
      <h2 class="title has-text-centered">BibTeX</h2>
      <span class="section-divider"></span>
      <div class="bibtex-block">
        <button class="bibtex-copy-btn" onclick="copyBibtex()" id="copyBtn">Copy</button>
<pre><code id="bibtexCode">@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}
}</code></pre>
      </div>
    </div>
  </section>

  <!-- ===== Footer ===== -->
  <footer class="site-footer">
    <div class="container has-text-centered">
      <p>
        This website is adapted from <a href="https://nerfies.github.io/">Nerfies</a> and <a href="https://mathvista.github.io/">MathVista</a>,
        licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0 International License</a>.
      </p>
    </div>
  </footer>

  <script>
  function copyBibtex() {
    var code = document.getElementById('bibtexCode').textContent;
    navigator.clipboard.writeText(code).then(function() {
      var btn = document.getElementById('copyBtn');
      btn.textContent = 'Copied!';
      setTimeout(function() { btn.textContent = 'Copy'; }, 2000);
    });
  }
  </script>

</body>
</html>