Research map / Image generation / Autoregressive and latent generation
EvoTok: A Unified Image Tokenizer via Residual Latent Evolution for Visual Understanding and Generation
Yan Li, Ning Liao, Xiangyu Zhao, Shaofeng Zhang, Xiaoxing Wang, Yifan Yang, Junchi Yan, Xue Yang
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 |
|---|---|---|---|
| use pixel-to-semantic evolution (Lpix=4, Lsem=16) instead of entangled equal depth (4,4) (ablation) | EvoTok · SEEDBench | accuracy: 62.7 -> 67.1 | Improved |
| use pixel-to-semantic evolution instead of entangled equal depth (ablation) | EvoTok · MME | score: 1668.9 -> 1793.5 | Improved |
| use pixel-to-semantic evolution instead of entangled equal depth (ablation) | EvoTok · GenEval | overall: 0.64 -> 0.67 | Improved |
| use pixel-to-semantic evolution instead of entangled equal depth (ablation) | EvoTok · ImageNet-1K val | rFID: 0.66 -> 0.55 | Improved |
| use pixel-to-semantic evolution instead of semantic-to-pixel (16,4) (ablation) | EvoTok · MME | score: 1731.6 -> 1793.5 | Improved |
| use pixel-to-semantic evolution instead of semantic-to-pixel (ablation) | EvoTok · GenEval | overall: 0.60 -> 0.67 | Improved |
| use pixel-to-semantic evolution instead of semantic-to-pixel (ablation) | EvoTok · ImageNet-1K val | rFID: 0.44 -> 0.55 | Hurt |
| use pixel-to-semantic evolution instead of semantic-to-pixel (ablation) | EvoTok · SEEDBench | accuracy: 64.6 -> 67.1 | Improved |
| use pixel-to-semantic evolution instead of layer-wise decoupled features (ablation) | EvoTok · ImageNet-1K val | rFID: 0.59 -> 0.55 | Improved |
| use shared-latent evolution instead of separate encoder branches (ablation) | EvoTok · ImageNet-1K val | rFID: 0.68 -> 0.55 | Improved |
| use EvoTok instead of UniTok trained on billion-scale data | EvoTok · ImageNet-1K val | rFID: 0.33 -> 0.43 | Hurt |
| use EvoTok instead of TokenFlow-B for understanding | EvoTok · MME | score: 1660.4 -> 1895.1 | Improved |
Closest papers by meaning
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.91)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.90)
- EvoComp: Learning Visual Token Compression for Multimodal Large Language Models via Semantic-Guided Evolutionary Labeling (CVPR 2026 · similarity 0.90)
- Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.90)
- G$^2$TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models (NeurIPS 2026 · similarity 0.89)
- UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation (ICLR 2026 · similarity 0.89)
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.
