[frakt0x90]

They created this in service of their video generation model which "clusters and reorders tokens based on semantic similarity using k-means.":

http://arxiv.org/pdf/2505.18875

[smusamashah]

Project website https://svg-project.github.io/