Research map / Image generation / Autoregressive and latent generation
SJD++: Improved Speculative Jacobi Decoding for Training-Free Acceleration of Discrete Auto-Regressive Text-to-Image Generation
Yao Teng, Zhihuan Jiang, Han Shi, Xihui Liu, Xuefei Ning, Guohao DaiMember, Yu Wang, Zhenguo Li, Xihui Liu
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 speculative Jacobi decoding with token reuse (SJD++) vs SJD (ablation) | Lumina-mGPT-7B · MS-COCO2017 val | latency acceleration: 2.04x -> 3.12x; step 2.22x -> 6.44x | Improved |
| add token reuse (SJD++) vs SJD (ablation) | Emu3-Gen-8B · MS-COCO2017 val | latency acceleration: 1.81x -> 3.01x; step 2.01x -> 7.50x | Improved |
| add token reuse (SJD++) vs SJD (ablation) | Janus-Pro-1B · MS-COCO2017 val | latency acceleration: 1.81x -> 2.31x | Improved |
| add token reuse (SJD++) vs SJD (ablation) | SimpleAR-0.5B · MS-COCO2017 val | latency acceleration: 1.94x -> 2.78x | Improved |
| add token reuse (SJD++) vs SJD (ablation) | Lumina-mGPT-7B · MS-COCO2017 val | FID: 30.90 -> 31.48 | No clear effect |
| use SJD++ instead of EAGLE or LANTERN or ZipAR | Lumina-mGPT · MS-COCO2017 val | latency acceleration: 2.56x -> 3.12x (LANTERN vs SJD++) | Improved |
| add Jacobi window size 96 instead of 16 (ablation) | SJD++ Janus-Pro-1B · Geneval | Geneval score: 0.7060 -> 0.7248 | Improved |
| use Jacobi window size 32 instead of 96 (ablation) | SJD++ Janus-Pro-1B · Geneval | Geneval score: 0.7248 -> 0.6838 | Hurt |
| disable refinement (replace unaccepted tokens with new initializations) (ablation) | SJD · image generation | forward passes: roughly one thousand -> more than two thousand | Hurt |
| use spatial-prior token initialization instead of random (ablation) | SJD · simple-pattern prompt | CLIP-Score: 31.37 -> 31.52 | No clear effect |
| increase resolution from 512 to 1024 (ablation) | SJD++ Lumina-mGPT-7B · Lumina-mGPT resolutions | step compression: up to 2.43x at 1024x1024 | Improved |
| use Jacobi decoding instead of SJD under stochastic top-K | Jacobi Decoding · COCO2017 Lumina-mGPT | latency acceleration: 1.02x vs 2.05x for SJD | No clear effect |
Closest papers by meaning
- Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding (ICLR 2025 · similarity 0.96)
- Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (NeurIPS 2025 · similarity 0.95)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.88)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.88)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.87)
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.
