D-ID (YC S17, [1]) promised that they would do the same. They have been quite silent on whether they ever achieved their target and nowadays they've pivoted to AI so no idea whether they ever actually succeeded.
It seems like the woman in the example is using a CV Dazzle makeup style and it doesn't fool the algorithm.
Other makeup styles work better, although I think that's probably just a short term solution (tweet is from 2018) before any new trendy makeup style is added to the training dataset.
https://adam.harvey.studio/cvdazzle