Artificially create more source images for Training.

Discussions about research, Faceswapping and things that don't fit in the other categories here.


Locked
User avatar
cosmico
Posts: 95
Joined: Sat Jan 18, 2020 6:32 pm
Has thanked: 13 times
Been thanked: 35 times

Artificially create more source images for Training.

Post by cosmico »

So I saw user tochans post about how you can get more source images,
and you should absolutely check out his perspective in the link below:
viewtopic.php?f=6&t=362
and it reminded me that I know of a way to artificially get additional angles of someones face from a single picture, and it might help out some of you guys if I shared it with y'all.

Its an app for your phone called "Muglife" and using some sort of 3d model technology, it allows you to play around with a face, make heads move angles to certain degree, and change facial emotions.
Image
This gif of the app is a totally-faked marketing trick, but it does represent what this app can do.

So check this out:
Let say you have some real "D" or even "F" list actor with very few pictures, the best you come up with is 20 good pics, for example this is one of those good pics below:
Image

Normally you'd be out of luck, why even bother trying to deepfake with just 20 images.
But with the app you can move their head angles around and even their emotions, and you might be able to turn
each one of those jpegs into a video where you can get 100 frames out of them.

Image
This mp4 file was shrunk down 40% and reduced to 25 fps when it was turned into a gif
Boom! You just turned 20 pics into 20: 100 frame videos, and now you have 2000 pics to train with.

It all sounds great right!? Well lets talk about the downsides:
-First and foremost, its cheap but it aint free. Massive bummer. You have to do like a 1.99 a week subscription, or like 30 a year.
-It doesn't 3-D move heads as far as you'd like them to get extra angles. It can probably go twice as far then what I showed in that gif, but it really started to look wonky, like when you stretch someones face too far. (but hey maybe faceswap wont care too much tho)
-It requires very good quality pics, that shouldn't be to big of a problem though hopefully.
-If the pic you give it already has its mouth open or is showing a lot of teeth in a smile (super common), it wont know how to read it right, and it can't properly animate the mouth without looking ridiculous. head and eyes will still work though.
-The results are great for everyday use, but when the guide recommends avoiding 1080p youtube, and instead using 4k, 2k, or ripping blu-rays, its not anywhere near that great of results. Depending on what phone you have and its pixel count, the maximum facial size you'll probably get will be around 1080x1920. I really have to stress this con. This trick will never be anywhere near as good as doing the deepfake the right way. You probably wont be making any masterpieces with this, but it may allow you actually do a deepfake that probably couldn't be done any other way. Or maybe just patching up a tiny metaphorical hole in your training set. Backup essentially.

User avatar
tochan
Posts: 21
Joined: Sun Sep 22, 2019 8:17 am
Been thanked: 5 times

Re: Artificially create more source images for Training.

Post by tochan »

At the moment, only on iOS... but... and now Android "now Early Access" i see!!! Now i can test the app on Andoid. Thanks for the reminder ;)

Locked