FlashAR speeds up pretrained autoregressive image models by 22.9x using 0.05% data

22.9倍加速！FlashAR：仅用0.05%数据，让预训练好的自回归图像模型飞起来

Zhejiang University and the University of Adelaide introduced FlashAR, using 0.05% of the original training data to reduce Emu3.5-Image-34B 512×512 generation latency from 130.10 seconds to 5.68 seconds, while GenEval changed from 80.48 to 80.29.

Why it matters: HKR-H/K/R all pass: FlashAR gives speedup, data ratio, latency, and GenEval deltas for AR image inference. It is a strong research item, but not a top-lab model release, so 80 featured rather than P1.