Research map / Image generation / Autoregressive and latent generation
LazyMAR: Accelerating Masked Autoregressive Models via Feature Caching
Feihong Yan, Qingyan Wei, Jiayi Tang, Jiajun Li, Yulin Wang, Xuming Hu, Huiqi Li, Linfeng Zhang
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 token cache (ablation) | MAR-H, 64 steps · ImageNet 256x256 class-conditional | speedup: 1.00 -> 1.59 | Improved |
| add token cache (ablation) | MAR-H, 64 steps · ImageNet 256x256 class-conditional | FID: 1.59 -> 1.62 | No clear effect |
| add condition cache (ablation) | MAR-H, 64 steps · ImageNet 256x256 class-conditional | speedup: 1.00 -> 1.87 | Improved |
| add condition cache (ablation) | MAR-H, 64 steps · ImageNet 256x256 class-conditional | FID: 1.59 -> 1.67 | No clear effect |
| add condition cache on top of token cache (ablation) | MAR-H, 64 steps · ImageNet 256x256 class-conditional | speedup: 1.59 -> 2.83 | Improved |
| add token cache on top of condition cache (ablation) | MAR-H, 64 steps · ImageNet 256x256 class-conditional | speedup: 1.87 -> 2.83 | Improved |
| select tokens by max similarity at 3rd layer instead of random caching (ablation) | LazyMAR, MAR-H 64 steps · ImageNet 256x256 class-conditional | FID: 1.95 -> 1.69 | Improved |
| select tokens by max similarity instead of min similarity (ablation) | LazyMAR, MAR-H 64 steps · ImageNet 256x256 class-conditional | FID: 2.10 -> 1.69 | Improved |
| cache decoded tokens instead of unpredicted tokens (ablation) | MAR-H 64 steps with condition cache · ImageNet 256x256 class-conditional | FID: 3.61 -> 1.79 | Improved |
| use similarity-based selection instead of caching only predicted tokens (ablation) | MAR-H 64 steps with condition cache · ImageNet 256x256 class-conditional | FID: 1.79 -> 1.69 | Improved |
| use LazyMAR instead of reducing decoding steps from 64 to 32 (similar speed) (ablation) | MAR-H · ImageNet 256x256 class-conditional | FID: 1.92 -> 1.69 | Improved |
| use LazyMAR instead of reducing decoding steps to 16 | MAR-B · ImageNet 256x256 class-conditional | FID: 4.10 -> 2.64 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MaskGIT — MaskGIT: Masked Generative Image Transformer (CVPR 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.90)
- CanvasMAR: Improving Masked Autoregressive Video Prediction With Canvas (NeurIPS 2026 · similarity 0.88)
- Look Ahead or Look Around? A Theoretical Comparison Between Autoregressive and Masked Pretraining (ICML 2024 · similarity 0.87)
- HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling (NeurIPS 2026 · similarity 0.87)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.87)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.87)
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.
