Research map / Image generation / Fast diffusion sampling
Denoising Token Prediction in Masked Autoregressive Models
Ting Yao, Yehao Li, Yingwei Pan, Zhaofan Qiu, Tao Mei
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 Transformer-based diffusion head (cross-attention over all tokens) (ablation) | MAR-B · ImageNet 256 (ablation, De-MAR-B, w/ CFG) | FID: 2.31 -> 2.10 | Improved |
| add denoising head with noised unmasked inputs (ablation) | MAR-B · ImageNet 256 (ablation, De-MAR-B, w/ CFG) | FID: 2.31 -> 2.22 | Improved |
| add denoising head on top of Transformer diffusion head (ablation) | MAR-B + Diff. Head · ImageNet 256 (ablation, De-MAR-B, w/ CFG) | FID: 2.10 -> 2.04 | Improved |
| add Transformer diffusion head on top of denoising head (ablation) | MAR-B + De. Head · ImageNet 256 (ablation, De-MAR-B, w/ CFG) | FID: 2.22 -> 2.04 | Improved |
| add Transformer diffusion head (ablation) | MAR-B · ImageNet 256 (ablation, De-MAR-B, w/ CFG) | IS: 281.7 -> 283.1 | Improved |
| add denoising head (ablation) | MAR-B · ImageNet 256 (ablation, De-MAR-B, w/ CFG) | Recall: 0.57 -> 0.57 | No clear effect |
| add both heads (De-MAR-B vs MAR-B) (ablation) | MAR-B · ImageNet 256 (ablation, De-MAR-B, w/ CFG) | Recall: 0.57 -> 0.59 | Improved |
| replace MAR with De-MAR (both heads) | MAR-B · ImageNet 256 (class-cond., w/o CFG) | FID: 3.48 -> 3.14 | Improved |
| replace MAR with De-MAR (both heads) | MAR-L · ImageNet 256 (class-cond., w/ CFG) | FID: 1.78 -> 1.66 | Improved |
| replace MAR with De-MAR (both heads) | MAR-H · ImageNet 256 (class-cond., w/ CFG) | FID: 1.55 -> 1.47 | Improved |
| replace MAR with De-MAR (both heads) | MAR · MS-COCO 256 text-to-image | FID: 6.36 -> 5.27 | Improved |
| scale De-MAR from Base to Huge | De-MAR · ImageNet 256 (class-cond., w/ CFG) | FID: 2.04 -> 1.66 -> 1.47 | Improved |
Closest papers by meaning
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.92)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.92)
- Learn from Your Mistakes: Self-Correcting Masked Diffusion Models (NeurIPS 2026 · similarity 0.90)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · similarity 0.90)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.90)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.90)
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.
