What is the quality gain when using a "Head" module in the new versions of RIFE?

[erikrause]

I have noticed, that the new versions of RIFE uses a "Head" nn.Module, which is not mentioned in paper and git source. Is there any ablation study on this module? It will be interesting to know about quality gain with this module. Also, what inspired you to introduce the "Head"?

[hzwer]

https://scholar.google.com/citations?view_op=view_citation&hl=zh-CN&user=ySHBXyAAAAAJ&citation_for_view=ySHBXyAAAAAJ:d1gkVwhDpl0C
The warp feature method has been around since some of the very early papers. My goal is to address the warp RGB bottleneck without introducing massive computational overhead.