How the style embedding for unseen subjects is handled during inference

[cavalier501]

Dear authors,

Thank you for your excellent work on FaceFormer. After carefully reading the main paper and supplementary material, I still have a question regarding how the style embedding for unseen subjects is handled during inference:

In the main paper (Section 4.1), you mention:

"For unseen subjects, we obtain the predictions of FaceFormer and VOCA by conditioning on all training identities. The implementation details of FaceFormer and the baseline methods are provided in the supplementary material (Sec. 1 and Sec. 2)."

However, the specific implementation details of "conditioning on all training identities" remain unclear to me:

1. Is any weighting strategy applied when conditioning on all training identities?
2. This issue is not explicitly addressed in Sections 1 and 2 of the supplementary material. Could you clarify the specific process used to handle unseen subjects?

Additionally, do you have insights on how similar problems are addressed in related works? I have reviewed some similar papers in this field, but most do not provide detailed explanations regarding this aspect.

Thank you in advance for your time and for sharing such impactful work! I look forward to your clarification.