Reproduce “Parallel Jacobi Decoding for Fast Autoregressive Image Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Boya Liao, Ying Li, Siyong Jian, Huan Wang
Parallel Jacobi Decoding (PJD) accelerates autoregressive image generation by leveraging the 2D spatial structure of images, enabling efficient parallel refinement across rows. This training-free method achieves up to 7.96x faster inference and 6.46x fewer decoding steps while maintaining high image quality compared to vanilla autoregressive models.
Key insights
- The 1D token expansion of traditional Jacobi decoding is suboptimal for images, which benefit more from 2D spatial parallelism due to strong local correlations.
- Dynamically activating rows based on contextual information from preceding rows maintains spatial coherence and reduces error propagation during parallel generation.
- A specially designed row-causal attention mask is crucial for enabling efficient parallel updates across rows while preserving the necessary autoregressive dependencies and preventing interference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.