AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

DualToken: Towards Unifying Visual Understanding and Generation with Dual Visual Vocabularies

Wei Song, Yuran Wang, Zijia Song, Yadong Li, Zenan Zhou, Long Chen, Xu Jhua, Jiaqi Wang, Kaicheng Yu

ICLR 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Vision-language pretraining

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train encoder for reconstruction only instead of text-aligned (SigLIP) pretraining (ablation)LLaVA-1.5 vision encoder · MMBenchaccuracy: 61.8 -> 36.2Hurt
apply recon and semantic losses on the same deep layer (single vocabulary) instead of hierarchical (ablation)DualToken tokenizer · ImageNet-1KrFID / zero-shot: rFID 0.25 -> 3.86; zero-shot 82.0 -> 72.3Hurt
apply reconstruction loss to shallow layer 6 and semantic loss to deep layer 26 (ablation)DualToken tokenizer · ImageNet-1KrFID: 3.86 -> 0.25Improved
add semantic loss to a reconstruction-only shallow tokenizer (ablation)DualToken tokenizer · ImageNet-1KrFID: 0.29 -> 0.25No clear effect
feed semantic plus pixel tokens instead of semantic tokens only (ablation)DualToken in LLaVA-1.5 · ten understanding benchmarksaverage score: 52.2 -> 53.9Improved
replace VILA-U tokenizer with DualToken in LLaVA-1.5LLaVA-1.5 · ten understanding benchmarksaverage score: 48.1 -> 53.9Improved
generate semantic tokens alongside pixel tokens instead of pixel tokens only (ablation)DualToken-3B · GenAI-Bench advancedoverall VQAScore: 0.55 -> 0.68Improved
use one unified tokenizer instead of separate MoVQGAN + SigLIP dual encoders (ablation)DualToken-3B · MJHQ-30KFID: 17.55 -> 7.88Improved
use DualToken 3B instead of VILA-U 3B tokenizerunified AR model · MJHQ-30KFID: 15.12 -> 7.88Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy