Research map / Image generation / Autoregressive and latent generation
Generation then Reconstruction: Accelerating Masked Autoregressive Models via Two-Stage Sampling
Feihong Yan, Yao Zhu, Peiru Wang, Pang Kaiyu, Qingyan Wei, 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 |
|---|---|---|---|
| use GtR sampling instead of original 64-step MAR | MAR-H · ImageNet 256x256 class-conditional | speedup / FID: 1.00 -> 3.72 speedup; FID 1.59 -> 1.59 | Improved |
| reduce sampling steps naively to 16 | MAR-H · ImageNet 256x256 class-conditional | FID: 1.59 -> 4.49 | Hurt |
| use GtR instead of LazyMAR at 8-step budget | MAR-H · ImageNet 256x256 class-conditional | FID: 4.06 -> 2.16 | Improved |
| use GtR sampling order instead of random order (ablation) | MAR-H · ImageNet 256x256 class-conditional | FID: 1.82 -> 1.59 | Improved |
| use GtR sampling order instead of raster order (ablation) | MAR-H · ImageNet 256x256 class-conditional | FID: 24.61 -> 1.59 | Improved |
| use GtR sampling order instead of 4-quadrant subsample (ablation) | MAR-H · ImageNet 256x256 class-conditional | FID: 5.19 -> 1.59 | Improved |
| add GtR to the encoder-decoder (ablation) | MAR-H · ImageNet 256x256 class-conditional | speedup: 1.43 -> 2.90 | Improved |
| add GtR to the diffusion head (ablation) | MAR-H · ImageNet 256x256 class-conditional | speedup: 1.43 -> 1.90 | Improved |
| add GtR to both encoder-decoder and diffusion (ablation) | MAR-H · ImageNet 256x256 class-conditional | speedup: 1.43 -> 3.73 | Improved |
| add Frequency-Weighted Token Selection (ablation) | GtR on MAR-H · ImageNet 256x256 class-conditional | FID: 1.65 -> 1.59 | Improved |
| select high-frequency tokens instead of random tokens for extra diffusion steps (ablation) | GtR on MAR-H · ImageNet 256x256 class-conditional | FID: 1.64 -> 1.59 | Improved |
| select high-frequency tokens instead of low-frequency tokens (ablation) | GtR on MAR-H · ImageNet 256x256 class-conditional | IS: 301.5 -> 304.4 | Improved |
Closest papers by meaning
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.93)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.92)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.91)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · 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.
