Research map / Vision-language and multimodal / Multimodal LLMs
WinTok: A Win-Win Hybrid Tokenizer via Decomposing Visual Understanding and Generation with Transferable Tokens
Yiwei Guo, Shaobin Zhuang, Zhipeng Huang, Canmiao Fu, Chen Li, Jing LYU, Yali Wang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| increase learnable tokens from 32 to 256 (ablation) | WinTok · ImageNet-1K | Top-1 accuracy: 78.8 -> 82.0 | Improved |
| increase learnable tokens from 32 to 256 (ablation) | WinTok · ImageNet-1K | rFID: 0.55 -> 0.42 | Improved |
| increase learnable tokens from 32 to 256 (ablation) | WinTok · MME-P | MME-P: 1248.0 -> 1453.0 | Improved |
| use SigLIP2-So400M teacher instead of CLIP-L/14 (ablation) | WinTok · MME-P | MME-P: 1328.4 -> 1453.0 | Improved |
| use SigLIP2-So400M teacher instead of DINOv2-L (ablation) | WinTok · MME-P | MME-P: 1274.5 -> 1453.0 | Improved |
| use SigLIP2 teacher instead of CLIP-L/14 (ablation) | WinTok · MMBench | accuracy: 58.1 -> 64.7 | Improved |
| use larger ViT-XL decoder instead of ViT-B (ablation) | WinTok · ImageNet-1K | rFID: 0.60 -> 0.54 | Improved |
| use larger ViT-XL decoder instead of ViT-B (ablation) | WinTok · generation | gFID: 16.51 -> 14.72 | Improved |
| swap roles of learnable and pixel tokens (LoseTok) (ablation) | WinTok · reconstruction | rFID | Hurt |
| add learnable semantic tokens with distillation vs UniTok | WinTok · ImageNet-1K | Top-1 accuracy: 70.8 -> 82.0 | Improved |
| use 256 instead of 64 semantic tokens in LLM (ablation) | WinTok · MMBench | accuracy: 60.2 -> 74.9 | Improved |
| use 256 instead of 64 semantic tokens in LLM (ablation) | WinTok · POPE | accuracy: 84.1 -> 86.5 | Improved |
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.
- UniTok — UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025) · better on 6 of 11 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
- UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation (ICLR 2026 · similarity 0.92)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.91)
- Rosetta Stone For Unified MLLMs: A Unified Tokenizer to Decipher Understanding and Generation (CVPR 2026 · similarity 0.90)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.88)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · similarity 0.88)
- Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.88)
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.
