Research map / Image generation / Autoregressive and latent generation
Latent Denoising Makes Good Tokenizers
Jiawei Yang, Tianhong Li, Lijie Fan, Yonglong Tian, Yue 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 |
|---|---|---|---|
| train tokenizer with latent denoising (l-DeTok) instead of MAR-VAE | MAR-B · ImageNet 256 | FID: 3.71 -> 2.43 | Improved |
| train tokenizer with latent denoising (l-DeTok) instead of MAR-VAE | RandomAR-B · ImageNet 256 | FID: 11.78 -> 5.22 | Improved |
| train tokenizer with latent denoising (l-DeTok) instead of MAR-VAE | RasterAR-B · ImageNet 256 | FID: 7.99 -> 4.46 | Improved |
| add semantics distillation (VA-VAE, MAETok) to tokenizers | MAR-B / RandomAR-B · ImageNet 256 | FID: VA-VAE 16.66 vs MAR-VAE 3.71 (MAR) | Hurt |
| use interpolative instead of additive latent noise (ablation) | tokenizer for SiT/MAR · ImageNet 256 | FID | Improved |
| add latent noising with tau~U(0,1) (ablation) | tokenizer MAR-B · ImageNet 256 | FID: 3.31 -> 2.77 | Improved |
| add latent noising with tau~U(0,1) (ablation) | tokenizer SiT-B · ImageNet 256 | FID: 6.97 -> 5.56 | Improved |
| shift noise level to higher values, logit(tau)~N(0.8,1) vs uniform (ablation) | tokenizer MAR-B · ImageNet 256 | FID: 2.77 -> 2.58 | Improved |
| use random instead of constant masking ratio (ablation) | tokenizer · ImageNet 256 | FID | Improved |
| add joint denoising (noise+masking) over baseline tokenizer (ablation) | tokenizer MAR-B · ImageNet 256 | FID: 3.31 -> 2.65 | Improved |
| add masking on top of latent noising (ablation) | tokenizer SiT-B · ImageNet 256 | FID | No clear effect |
| scale generative model with l-DeTok vs baseline tokenizer | SiT-XL · ImageNet 256 | FID: 4.47 -> 3.14 | Improved |
Closest papers by meaning
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.91)
- SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer (CVPR 2025 · similarity 0.89)
- Epsilon-VAE: Denoising as Visual Decoding (ICML 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.88)
- Highly Compressed Tokenizer Can Generate Without Training (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.
