This AI Turns Pixel Faces Into ‘Photos’
Face Depixelizer is an amazing new AI-powered app that can take an ultra-low-res pixelated photo of a face and turn it into a realistic portrait photo.
After taking in a pixel face, Face Depixelizer continuously creates faces with StyleGAN and narrows in on the resulting “photo” by finding one that downscales to produce the exact same pixel face.
Here’s a short animation showing the process in action:
The input faces don’t even need to be real people — you can use the tool to see what video game characters might look like in real life. Here’s what Malimonov got when he put in Wolfenstein protagonist B.J. Blazkowicz:
With default settings, I got this result. pic.twitter.com/mRkqqTwhJF
— Bomze (@tg_bomze) June 20, 2020
And here’s what you get with Doom:
So that one is Wolfenstein, here is Doom pic.twitter.com/ixpItQyJ5q
— Hani ✊🏽🧢 (@h_bash) June 20, 2020
As with all AI systems of this sort, though, your mileage may vary:
All interpretations downscale to the same pixels, so this is an equally valid real life DoomGuy. pic.twitter.com/RwjZSFk9Tg
— Jonathan Fly 👾 (@jonathanfly) June 20, 2020
— Moritz Klack (@moklick) June 20, 2020
It seems that one current issue with the system — perhaps due to the faces StyleGAN was trained on — is that it turns black people into white people.
*beep boop* AI does blackface. pic.twitter.com/BeUDoAsfvj
— Kiloku (@Kiloku) June 20, 2020
🤔🤔🤔 pic.twitter.com/LG2cimkCFm
— Chicken3gg (@Chicken3gg) June 20, 2020
Here’s what we got a result after inputting a pixelized photo of Leonardo DiCaprio:
Some people are pointing out that this type of tool could be used to identify faces that have intentionally been pixelated for privacy or safety, but Malimonov’s response is that the tool doesn’t show what the person actually looks like — it just finds any face that fits.
This tool will not restore the original face, however, it can help with the identification of facial features.
— Bomze (@tg_bomze) June 20, 2020
You can use Face Depixelizer yourself over at Google Colab. The input photo needs to be square, and you get started by pressing the button indicated by the red arrow and then scrolling down to where the browse file button appears.
A big thanks to Sam Cornwell of Solarcan for the tip!
