A curated list of papers, methods, and resources on accelerating autoregressive (AR) image generation.
π₯ Pull Requests are welcome! Feel free to contribute new papers or methods.
Autoregressive (AR) image generation has shown strong performance in visual synthesis tasks, but its sequential decoding nature leads to slow inference.
This repository collects research papers and methods aiming to accelerate AR image generation, including:
- π Faster decoding algorithms
- Accelerating Auto-Regressive Text-to-Image Generation with Training-Free Speculative Jacobi Decoding.ICLR,2025 [Paper] [Code]
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding.ICLR,2025 [Paper] [Code]
- Grouped Speculative Decoding for Autoregressive Image Generation. ICCV,2025 [Paper] [Code]
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality. ICLR,2025 [Paper] [Code]
This repository follows the MIT License unless otherwise noted.
If you find this repository useful, please β star it!