Research map / Image generation / Autoregressive and latent generation
CaTok: Taming Mean Flows for One-Dimensional Causal Image Tokenization
Yitong Chen, Zuxuan Wu, Xipeng Qiu, Yu-Gang Jiang
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 |
|---|---|---|---|
| add MeanFlow loss to rectified-flow decoder (one-step rFID) (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 ablation (80 epochs) | rFID@1: 183.69 -> 4.71 | Improved |
| add MeanFlow loss (generation) (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 ablation (80 epochs) | gFID: 19.67 -> 24.39 | Hurt |
| add REPA objective (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 ablation (80 epochs) | gFID: 24.39 -> 17.92 | Improved |
| add REPA-A objective on encoder features (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 ablation (80 epochs) | gFID: 17.92 -> 13.54 | Improved |
| add token selection in interval [r,t] (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 ablation (80 epochs) | gFID: 13.54 -> 4.91 | Improved |
| add token selection in [r,t] (one-step reconstruction) (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 ablation (80 epochs) | rFID@1: 3.92 -> 4.89 | Hurt |
| select tokens in [r,t] instead of all tokens (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 token selection | gFID: 13.54 -> 4.91 | Improved |
| select tokens in [r,t] instead of first k tokens (ablation) | CaTok-B-256 · ImageNet-1K 256, CaTok-B-256 token selection | gFID: 9.21 -> 4.91 | Improved |
| add MeanFlow-style causality (with token selection) in non-autoregressive generator (ablation) | CaTok-L-32 · ImageNet-1K 256, eps-MaskGiT 8 steps | gFID: 3.26 -> 2.69 | Improved |
| use CaTok tokenizer instead of LlamaGen tokenizer (VQ) | LlamaGen generator · ImageNet-1K 256 (CaTok-VQ with FSQ) | gFID: 3.80 -> 3.35 | Improved |
| use CaTok-L-256 instead of Semanticist-L-256 without CFG | AR generator · ImageNet-1K 256 generation, no CFG | gFID: 7.60 -> 5.52 | Improved |
| increase number of tokens from 32 to 128 | CaTok-L · ImageNet-1K 256 generation | gFID: 3.40 -> 2.95 | Improved |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.88)
- LARP: Tokenizing Videos with a Learned Autoregressive Generative Prior (ICLR 2025 · similarity 0.88)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.88)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.88)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 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.
