Shaving Henry Cavill’s mustache: a DeepFakes VFX demo
If youg didn’t hear about it, Henry Cavill’s mustache became quite famous during MustacheGate. Cavill grew a mustache for his role in the upcoming Mission Impossible 6. Later, the Justice League film needed last-minute reshoots. The simple thing to do would be to shave the mustache for Justice League and use a glue-on prop for MI6, but the competing MI6 studio would not allow that. As a result, Cavill’s mustache was digitally removed and forever enshrined as one of the worst examples of facial hair VFX to ever grace the silver screen. Even the Telegraph reported on the mustache antics.
Justice League had a budget of $300 million. I used a $500 computer bought off Craigslist and free deepfakes software to teach an AI how to digitally remove Henry Cavil’s mustache in a variety of situations. I also attempted to restore Superman’s correct face. You can see the side-by-side comparisons in the video below.
While deepfaking usually involves swapping two faces (see our Nicolas Cage post for an example), I instead trained the computer model on two different version, mustached and shaven, of the same face. The model did a decent job of removing Cavill’s mustache, at least as cleanly as the average rushed morning shaving session. Due to the inflexibility of the current blending functions in the deepfaking software, I saved the shaved images without any blurring or blending functions. I then masked the shaved mouths and combined them with the original clips in video editing software.
At the end of the clip, you will see an attempt at a reverse edit, in which I add a mustache to a cleanly shaven Henry Cavill. Adding a mustache was less robust, in general, then removing the mustache. I had difficulty finding diverse training data sets for the mustached Henry Cavill, and that may have played a role. There are just more diverse images of Henry Cavil without his mustache.
This post is the first in a series of planned demos to showcase VFX techniques made accessible by deepfaking.