Research map / Image generation / Autoregressive and latent generation
Speculative Coupled Decoding for Training-Free Lossless Acceleration of Autoregressive Visual Generation
Junhyuk So, Hyunho Kook, Chaeyeon Jang, Eunhyeok Park
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 |
|---|---|---|---|
| replace independent draft sampling with maximal coupling (pi_MC) in SJD (ablation) | SJD on Lumina-mGPT · MS-COCO, L=64 | NFE: 1035.9 -> 567.7 | Improved |
| replace independent draft sampling with Gumbel coupling (pi_GS) in SJD (ablation) | SJD on Lumina-mGPT · MS-COCO, L=32 | NFE: 1031.2 -> 652.3 | Improved |
| increase window size L with plain SJD (ablation) | SJD on Lumina-mGPT · MS-COCO | NFE: 1058.6 (L=16) -> 1035.9 (L=64) | No clear effect |
| increase window size L with coupling (ablation) | SCD (pi_MC) on Lumina-mGPT · MS-COCO | NFE: 814.5 -> 567.7 | Improved |
| add coupling on top of SJD (ablation) | SJD on Lumina-mGPT · MS-COCO | FID: 30.81 -> 30.83 | No clear effect |
| use lossless coupling instead of lossy grouped speculative decoding (GSD) | Lumina-mGPT · MS-COCO | FID: 33.21 -> 30.75 | Improved |
| add coupling on top of SJD for Janus-Pro | SJD on Janus-Pro 7B · MS-COCO, L=32 | NFE: 318.01 -> 154.42 | Improved |
| add coupling on top of SJD for video AR | SJD on Cosmos-1-AR-4B · Real-Estate-10k, L=128 | latency (s): 47.73 -> 13.60 | Improved |
| increase coupling strength alpha from 0 to 1 (ablation) | SJD on Lumina-mGPT · NFE vs mean token difference | NFE | Improved |
| use pi_GS instead of pi_MC for easy-to-draft tasks (video, low-resolution images) (ablation) | SCD · video AR and Janus-Pro | NFE: 577.8 -> 564.4 (video L=128) | Improved |
| increase CFG scale (ablation) | SCD · Lumina-mGPT | speedup | Hurt |
Closest papers by meaning
- Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding (ICLR 2025 · similarity 0.92)
- Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (NeurIPS 2025 · similarity 0.92)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.91)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.90)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.90)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (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.
