Token merging has emerged as a practical, training-free way to speed up diffusion models by cutting down on redundant token computation. The difficulty lies in deciding which tokens can be safely merged without noticeably degrading the generated image. A new paper on arXiv tackles this selection problem directly, proposing a method that uses diffusion dynamics to guide the merging decisions.

According to the abstract, the approach aims to reduce computation while keeping generation quality intact. Because the method is training-free, it can be applied to existing models without additional optimization. The paper's title emphasizes the core question—"Which Tokens to Merge?"—and positions diffusion dynamics as the answer.

This summary is based solely on the provided abstract, which is truncated. No comparison with other sources is possible, and specific experimental results are not detailed in the available text.