Research map / Image generation / Autoregressive and latent generation
MacTok: Robust Continuous Tokenization for Image Generation
Hengyu Zeng, Xin Gao, Guanghao Li, Yuxiang Yan, Jiaoyang Ruan, Junpeng Ma, Haoyu Albert Wang, Jian Pu
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 local representation alignment to random masking (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B) | gFID: 6.01 -> 3.53 | Improved |
| add DINO-guided semantic masking (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B) | gFID: 3.53 -> 3.32 | Improved |
| add global representation alignment (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B) | gFID: 3.32 -> 3.15 | Improved |
| use equal mix of random and semantic masking instead of random only (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B, w/o CFG) | gFID: 14.59 -> 13.95 | Improved |
| use 100% semantic masking instead of random only (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B, w/o CFG) | gFID: 14.59 -> 14.84 | No clear effect |
| raise max mask ratio from 0.4 to 0.7 (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B, w/o CFG) | gFID: 15.49 -> 14.59 | Improved |
| raise max mask ratio from 0.7 to 0.8 (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B, w/o CFG) | gFID: 14.59 -> 14.92 | Hurt |
| add decoder fine-tuning (ablation) | MacTok-128 · ImageNet 256x256 (SiT-B, w/o CFG) | rFID: 0.57 -> 0.43 | Improved |
| use MacTok tokenizer instead of SoftVQ-VAE at 64 tokens | SiT-B · ImageNet 256x256 (optimal CFG) | gFID: 4.09 -> 3.22 | Improved |
| increase MacTok tokens from 64 to 128 | MacTok+SiT-XL · ImageNet 512x512 (w/o CFG) | gFID: 4.63 -> 5.12 | Hurt |
Closest papers by meaning
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.90)
- FlexTok: Resampling Images into 1D Token Sequences of Flexible Length (ICML 2025 · similarity 0.89)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.88)
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.88)
- SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer (CVPR 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.
