Why is the energy first normalized in each frame before performing top-k selection across all frames?

[ThisisBillhe]

Thank you for your great work!

I have a question regarding the top-k energy selection process. Since the energy is normalized per frame, the values across different frames are on different scales. Could you please explain the rationale behind performing a global top-k selection across all frames after this per-frame normalization?

[YunyaoYan]

Thanks for the insightful question!

We’ve prepared a short explainer (with a simple visual demo) to illustrate why we use per-frame normalization before global Top-K:
👉 https://evolvinglmms-lab.github.io/OneVision-Encoder/issues/issue-116.html

Let us know if you have any follow-up questions!