Research map / Image generation / Autoregressive and latent generation
MARché: Fast Masked Autoregressive Image Generation with Cache-Aware Attention
Chaoyi Jiang, Sungwoo Kim, Lei Gao, Hossein Zarch, Won Woo Ro, Murali Annavaram
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 caching tokens to active set (ablation) | MARché · ImageNet 256 ablation | FID: 2.70 -> 2.56 | Improved |
| add generating tokens to active set (ablation) | MARché · ImageNet 256 ablation | FID: 504.82 -> 2.56 | Improved |
| use attention-guided selection instead of random token recomputation (ablation) | MARché · ImageNet 256 ablation | FID: 564.61 -> 2.56 | Improved |
| select refreshing tokens by high attention instead of low attention (ablation) | MARché · ImageNet 256 ablation | FID: 3.80 -> 2.56 | Improved |
| select refreshing tokens by high attention instead of random (ablation) | MARché · ImageNet 256 ablation | FID: 3.01 -> 2.56 | Improved |
| add refreshing tokens (64) to caching (ablation) | MARché · ImageNet 256 ablation | FID: 3.17 -> 2.56 | Improved |
| use layer 4 instead of layer 1 for refresh token selection (ablation) | MARché · ImageNet 256 ablation | FID: 2.62 -> 2.49; latency 0.154s -> 0.1613s | Mixed |
| full refresh at layers 1-6 instead of layer 1 only (ablation) | MARché · ImageNet 256 ablation | FID: 2.62 -> 2.49; latency 0.155 -> 0.164 | Mixed |
| full refresh at layers 7-12 instead of layer 1 only (ablation) | MARché · ImageNet 256 ablation | FID: 2.62 -> 25.35 | Hurt |
| full refresh on even layers instead of layer 1 only (ablation) | MARché · ImageNet 256 ablation | FID: 2.62 -> 4.65 | Hurt |
| replace concatenated attention with cache-aware attention kernel (ablation) | MARché · kernel latency (lengths 32 and 256) | latency: 904.2 us -> 730.3 us | Improved |
| apply cache-aware inference to MAR-B | MAR-B · ImageNet 256x256 | FID: 2.35 -> 2.56; latency 0.104 -> 0.064 | Hurt |
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 2 of 2 shared results
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 0.91)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.89)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.89)
- HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling (NeurIPS 2026 · similarity 0.88)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (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.
