Research map / Image generation / Autoregressive and latent generation
LISA: Locality-Informed Speculative Decoding for Accelerating Autoregressive Image Generation
Ying Li, Siyong Jian, Zhaode Wang, Zhiwen Chen, Chengfei Lyu, Huan Wang
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 geometry-aware soft verification (GASV) (ablation) | LISA drafter · LlamaGen-XL-T2I | accept length: 1.24 -> 2.92 | Improved |
| add GASV (ablation) | LISA drafter · LlamaGen-XL-C2I | accept length: 1.07 -> 2.45 | Improved |
| add GASV (ablation) | EAGLE-3 drafter · LlamaGen-XL-T2I | accept length: 1.00 -> 1.32 | Improved |
| add GASV (ablation) | EAGLE-3 drafter · LlamaGen-XL-C2I | accept length: 1.00 -> 1.25 | Improved |
| use locality-informed distilled drafter instead of EAGLE-3 (standard verification) (ablation) | LISA · LlamaGen-XL-T2I | accept length: 1.00 -> 1.24 | Improved |
| use locality-informed distilled drafter instead of EAGLE-3 (standard verification) (ablation) | LISA · LlamaGen-XL-C2I | accept length: 1.00 -> 1.07 | Improved |
| increase drift threshold kappa from 0.6 to 1.0 (ablation) | LISA · LlamaGen-XL-T2I | accept length: 2.16 -> 2.92 | Improved |
| increase cluster-posterior threshold lambda from 0.4 to 0.8 (ablation) | LISA · LlamaGen-XL-T2I | accept length: 2.92 -> 2.49 | Hurt |
| loosen verification (kappa=1.1, lambda=0.4 vs kappa=1.0, lambda=0.6) (ablation) | LISA · LlamaGen-XL-C2I | speedup and FID: 1.39x -> 1.53x speedup; FID 10.50 -> 12.60 | Mixed |
| apply locality-informed distillation to Medusa-style drafter (ablation) | Medusa-style drafter · LlamaGen-XL-T2I | speedup: 1.04x -> 1.37x | Improved |
| add GASV to LISA-Medusa (ablation) | LISA-Medusa · LlamaGen-XL-T2I | speedup: 1.37x -> 1.82x | Improved |
| use LISA instead of EAGLE-3 | Lumina-mGPT-7B · MS-COCO | speedup: 1.86x -> 4.12x | 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.
- ZipAR-16 — ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025) · 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
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.92)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.91)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.91)
- Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 2025 · similarity 0.89)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.89)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (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.
