SSD: A New Method to Accelerate Image Generation
rohanpaul_ai · x · 2026-07-14
This paper introduces Spatially Speculative Decoding (SSD), which uses spatial parallel prediction to accelerate autoregressive image generation, achieving speedups of up to 13.28×.
Method Highlights
- Traditional image autoregressive models convert 2D images into long token sequences and generate them token by token
- Instead of guessing single tokens, SSD uses a small helper network leveraging spatial information from the right and bottom to predict entire rows or blocks simultaneously
- These helper networks predict internal features that are easier to learn, rather than final discrete labels
- The main model verifies the entire block draft at once, correcting errors and reducing wasted computation
Experimental Results
- Achieved a 5.74× to 13.28× reduction in generation time on Janus-Pro, Lumina-mGPT, and Emu3
- Benchmark scores remained largely on par with standard decoding
- Some of these gains come from a lightweight inference loop, not just from draft parallelism itself
Conclusion
The authors believe that larger token grids will yield more significant speedup benefits, and this approach can drastically reduce the wall-clock time for image generation without noticeably sacrificing performance.
More from Research
- Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
- RSS launches under OMSF to push structural biology data modeling at scale — MoAlQuraishi · 2026-07-22
- enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
- enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22
- enFoldX gains accuracy as AF3 ensemble disagreement rises for non-binders — quaidmorris · 2026-07-22
- A 3D ray plot shows how hard this Jacobian counterexample is to read — moultano · 2026-07-22