Research map / Image generation / Autoregressive and latent generation
Annealed Relaxation of Speculative Decoding for Faster Autoregressive Image Generation
Xingyao Li, Fengzhuo Zhang, Cunxiao Du, Hui Ji
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 annealed per-step relaxation (Cool-SD) instead of uniform relaxation parameter (ablation) | relaxed speculative decoding · LlamaGen-XL and Lumina-mGPT, MS-COCO val | FID vs accepted length/latency | Improved |
| replace LANTERN++ resampling distribution with the derived optimal G* (ablation) | LANTERN++ · LlamaGen-XL and Lumina-mGPT, MS-COCO val | FID vs speed trade-off | Improved |
| larger k in LANTERN++ with its original resampling (ablation) | LANTERN++ · MS-COCO val | FID | Hurt |
| add Cool-SD relaxation (delta=1.1) over lossless Eagle-1 | Eagle-1 · LlamaGen-XL, MS-COCO val | accepted length: 2.42 -> 2.73 | Improved |
| raise relaxation budget to delta=2 | Cool-SD · LlamaGen-XL, MS-COCO val | accepted length: 2.42 -> 3.34 | Improved |
| increase delta from 1.1 to 2 (ablation) | Cool-SD · LlamaGen-XL, MS-COCO val | FID: 21.02 -> 21.20 | Mixed |
| add Cool-SD relaxation over Eagle-1 | Eagle-1 · Lumina-mGPT, MS-COCO val | accepted length: 2.75 -> 3.15 | Improved |
| Cool-SD instead of LANTERN++ at similar quality | LANTERN++ · Lumina-mGPT, MS-COCO val | speed-up: 2.48 -> 2.69 | Improved |
| Cool-SD instead of Speculative Jacobi Decoding | SJD · Lumina-mGPT, MS-COCO val | speed-up: 2.06 -> 2.69 | Improved |
| Cool-SD instead of Speculative Jacobi Decoding | SJD · LlamaGen-XL, MS-COCO val | speed-up: 1.84 -> 2.27 | Improved |
Closest papers by meaning
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.92)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.91)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- A Theoretical Perspective for Speculative Decoding Algorithm (NeurIPS 2024 · similarity 0.89)
- Speculative Speculative Decoding (ICLR 2026 · similarity 0.89)
- Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (NeurIPS 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.
