Skip to content

FlashGaze

Ziye Zhu1,2Yanghao Zhou2Lixing Tan2 Jialiang Kang2Shuxuan Li2,3Xiao Yang2,*

1 2 3 *

98%

13×

1.8×

1.5K/4K

At* = argminAt [ ∑(p,a)∈At Cost(p, a) + λNt ] s.t. Nt ≤ B

@misc{zhu2026flashgaze,
  title        = {{FlashGaze}: Training-Free Multi-Scale Patch Pruning for Efficient Video Understanding},
  author       = {Zhu, Ziye and Zhou, Yanghao and Tan, Lixing and Kang, Jialiang and Li, Shuxuan and Yang, Xiao},
  howpublished = {arXiv preprint arXiv:2610.04225},
  year         = {2026}
}