Research map / Image generation / Autoregressive and latent generation
Autoregressive Image Generation Needs Only a Few Lines of Cached Tokens
Ziran Qin, Youru Lv, Mingbao Lin, Zeren Zhang, chaofan gan, Tieyuan Chen, Liquan Shen, Junhui Hou, Chern Hong Lim, Fei Wen, Weiyao Lin
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 initial anchor tokens to the cache (ablation) | LineAR (Janus-Pro-1B, rho=1/6) · MS-COCO | FID-30k: 203.38 -> 22.86 | Improved |
| add recent lines cache (ablation) | LineAR (Janus-Pro-1B, rho=1/6) · MS-COCO | FID-30k: 79.96 -> 22.86 | Improved |
| add inter-line selected mid-region tokens (ablation) | LineAR (Janus-Pro-1B, rho=1/6) · MS-COCO | FID-30k: 24.80 -> 22.86 | Improved |
| use inter-line guidance instead of attention accumulation (rho=1/6) (ablation) | LineAR (Janus-Pro-1B) · MS-COCO | FID-30k: 23.27 -> 22.86 | Improved |
| use inter-line guidance instead of attention accumulation (rho=1/8) (ablation) | LineAR (Janus-Pro-1B) · MS-COCO | FID-30k: 24.30 -> 24.03 | Improved |
| use inter-line guidance on GenEval (rho=1/6) (ablation) | LineAR (Janus-Pro-1B) · GenEval | GenEval overall: 0.68 -> 0.71 | Improved |
| apply LineAR at 1/6 budget | LlamaGen-XL · ImageNet class-conditional | FID-50k: 2.77 -> 2.68 | Improved |
| apply LineAR at 1/6 budget | Janus-Pro-1B · MS-COCO | FID-30k: 23.85 -> 22.86 | Improved |
| apply LineAR at 1/6 budget | Janus-Pro-1B/7B · GenEval and DPG | GenEval overall: 0.73 -> 0.71 | Hurt |
| apply LineAR at 1/8 budget | Lumina-mGPT-768 · DPG | DPG overall: 76.96 -> 77.19 | Improved |
| reduce budget ratio from 1/4 to 1/8 | LineAR (LlamaGen-XL) · ImageNet class-conditional | FID-50k: 2.64 -> 2.95 | Hurt |
| use LineAR instead of R-KV at 1/8 budget | LlamaGen-XL · ImageNet class-conditional | FID-50k: 3.45 -> 2.95 | Improved |
Closest papers by meaning
- HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling (NeurIPS 2026 · similarity 0.90)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.90)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.89)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.89)
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.88)
- Parallelized Autoregressive Visual 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.
