This AI Learned to “Photoshop” Human Faces

Károly Zsolnai-Fehér explains a neural network that edits faces photorealistically, changing jawlines, smiles, accessories, hair color, and makeup. It processes images in 50 milliseconds and the authors released the source code openly.
Questions this video answers
March 27, 2019.
Roughly 50 milliseconds (≈20 images per second).
512×512.
Yes — available free of charge and under a permissive license.
Károly Zsolnai-Fehér.
Chapters
Dear fellow scholars, this is two minute papers with.
In this series we talk quite a.
Bit about neural network based learning methods.
That are able to generate new images for us from some sort of sparse description like a written sentence or a set of controllable parameters. These can enable us mere mortals without artistic skills to come up with novel images. However, one thing that comes up with.
Almost every single one of these techniques.
Is, is the lack of artistic control. You see, if we provide a very coarse input, there are many, many different.
Ways for the neural networks to create.
Photorealistic images from them. So how do we get more control over these results? An earlier paper from Nvidia generated human.
Faces for us and used a latent.
Space technique that allows us some more.
Fine grained control over the images.
It is beyond amazing, but these are called latent variables because they represent the inner working process of the neural network and they do not exactly map to our intuition of facial features in reality.
And now have a look at this.
New technique that allows us to edit.
The geometry of the jawline of a person, put a smile on someone's face.
In a more peaceful way than seen.
In some Batman movies, or remove the.
Sunglasses and add some crazy hair at the same time.
Even changing the hair of someone while adding an earring with a prescribed shape is also possible. Whoa. And I just keep talking and talking about artistic control.
So it's great that these shapes are.
Supported, but what about another important aspect of artistic control?
For instance, colors?
Yep, that is also supported here. You can see that the color of the woman's eyes can be changed and.
And the technique also understands the concept of makeup as well. How cool is that?
Not only that, but it is also blazing fast.
It takes roughly 50 milliseconds to create.
These images with the resolution of 512x512. So in short, we can do this about 20 times per second. Make sure to have a look at.
The paper that also contains a validation.
Section against other techniques and reference results. Turns out there is such a thing.
As a reference result in this case.
Which is really cool.
And you will also find a novel.
Style loss formulation that makes all this crazy wizardry happen.
No web app for this one.
However, the source code is available free of charge and under a permissive license.
So, let the experiments begin.
If you have enjoyed this video and.
You feel that a bunch of these.
Videos are worth $3 a month, please.
Consider supporting us on Patreon. In return, we can offer you early access to these episodes and more to.
Keep your paper addiction in check.
It is truly a privilege for me.
To be able to keep making these videos.
I am really enjoying the journey and.
This is only possible because of your support on Patreon. This is why every episode ends with.
You guessed it right. Thanks for watching and for your generous.
Support, and I'll see you next time.










