We Break FaceApp to See How It Works

We Break FaceApp to See How It Works

[Sounds of shock and disgust] Niko: If I want to make FaceApp basically what I do … [Old Niko] Hehehe. … just draw some eyes on a whiteboard and see where it applies the detail. Niko: So obviously my Instagram feed just blowing up with pictures of old people. Turns out it’s all my friends just using FaceApp. Nick: We made a video about FaceApp like a year ago and the results were pretty hilarious. Niko: I’ve ever been doing a ton of work with AI face manipulation. Particularly our deep fake videos. Now I think the technology underneath the program is actually pretty similar. I want to see what happens when we apply it to our pano’s. [Sam’s Death Squeal] We’re gonna take pictures of people that we have, from both being young and old, and we’re gonna compare to see how accurate it is. But I also want to figure out how it works. So we’re gonna try to uncover that mystery for you today. Clint: Is this kind of awkward that I’m wearing a suit? Niko: You and me, Wren, we’re buddies, out on a retirement cruise. (Woah!) Clint: They even got the neck rolls. Nick: It’s me. [Laughter] Clint: Oh my god, dude. They do the hand too. Nick: Eww! Clint: Looks like a freaking camel spider bite. Nick: It was the war. The war got me. The Great War of 2030. Niko: Alright Jake. This is … Jake: … exactly the same. Damn son. Niko: He’s a guy we named Bob. Bob Watson. Nick: He’s like a roof shingler. He lets himself inside your kitchen to get like a Coors Light from the fridge. Jake: Bob stop drinking on the job again, Bob. Nick, are you just insinuating that I’m just gonna be an old drunk? [Laughter] Clint: Look at this. Twice old Jean- Jimmy Carter, dude. Jimmy Carter! Clint: Got him. Dean: Got ’em! Niko: [Laughter] It’s like super super accurate. Like here … Wren put your hand up next to your face. It’s doing your wrist wrinkles correctly and your fingers correctly. Nick: What is this giggle person? [Sam’s Death Squeal] Clint: I’m going to see if it works on a pano. [Sam’s Death Squeal] Wren: [Reacts] Hohh! Heuuuu … Nick: The teeth are pretty good. Clint: Stage 1 … Stage 2 … … Stage 3. Wren: Dude that’s like an Attack on Titan. Clint: I know I know. [Sam’s Death Squeal] Clint: Alright. Are you guys ready for stage 4 though? Nick: Four levels deep man…? Oh no. Oh no-hohoho. – Dude! – Dude, it’s a titan. – The neck … [Laughter] Clint: Hang on. Let’s de-age. Let’s de-age. [Yells of shock] [Laughter] Niko: So the way like it’s kind of patterning the wrinkles It kind of has that feeling like it’s basically just learned the little patterns from other images And it’s kind of recognizing where those patterns will fit in on this image. I’m … suspicious that this is actually using machine learning and basically the same thing that we’re doing with deep fakes Wren: I say 100%. This is *absolutely* machine learning. It’s not just aging up the face. It’s like it’s actually adding hair where necessary. It’s adding like neck wrinkles down here and like you’re saying it’s even adding wrinkles to the skin on the arms and various other aspects. Pretty much anything that looks youthful, it’s changing. Niko: Let’s do a quick accuracy test here. Niko: Here we go … wow! That’s pretty good. Dean: It looks … exactly like Ian McKellen, honestly. Niko: Let me take a picture of a young celebrity and I’ll show you the old version and you tell me who the celebrity is. Okay, so step away for a second. Jake: The face is right here, Niko, if you need a young celebrity. Niko: Alright. Don’t look. Don’t look. Niko: [Stealthily] Hiding it so Jake can’t see it. [Multiple] Wow! Woooaaah! Niko: It’s pretty obvious. [Pretend ignorant voice] Jake who’s this celebrity? Jake: Uhhhhhhh … It’s Robert De Niro, obviously. [Niko Laughing] Niko: Christian. You don’t know who I took a picture of when they were young but tell me who is this? Christian: Morgan Freeman. Dean: Yeah, obviously. Christian: If you showed me the younger one it might have been more confusing. Niko: Slightly more challenging one. Who is this? Wren: Oh! It’s the king’s dude … the Kingsman. Colin Firth! That’s his name. Niko: Yep. [Laughter] Wren: It’s becoming really apparent that everyone kind of ages the same way. We all have like skin that starts sagging in the same spots. We all have wrinkles that form in the same spots. I really wanna see … a video with this and I think the way to do that is basically just take the video. Render it out as an image sequence. Import all those images into your phone. Then you have to manually go through and do it. I really wonder how it’s gonna change between each frame. There’s no continuity between before and after. I’m really curious to see how it goes over time. [Spongebob voice] Tomorrow. Dean: Hey Niko. Niko: What am I about to see? [Old man Niko voice] Heheheh. I’m an old man. Hurhurhur. [inhales] Niko: That actually looks pretty good. The wrinkles stay really consistent on my face. The beards a little all over the place. So yeah, this is really impressive. They could probably make it temporally consistent. You know what? I bet that that is one year away. I bet you within one year, we’ll see FaceApp for video. 61.4 percent of our watch time is from people who aren’t subscribed. Only 34 percent of the people. So if you’re not subscribed, please consider doing so. It helps out the videos. Helps out the channel. [Disembodied Niko voice] Subscribe! Subscribe! [Rewinding sound] Niko: Think we can nail down exactly what it’s doing. You know what we could also try is a fake face. Like just draw some eyes on a whiteboard and like see where it applies the detail. So what we need to do, is we need to recognize: how is it applying these wrinkles? How is it learning this texture and applying it and I think we can kind of pinpoint some patterns. I think we can nail down exactly what it’s doing. Dean: Oooh, is it going to see if there’s a face there? Niko: Ohh. It did. It did! Nick: What’s it going to do? So it seems like it’s working. So let’s old it I guess. Niko: Make that face old. [Wren laughter] Wren: It just added wrinkles to the whiteboard. Nick: But that’s like perfect wrinkle, like, test right there. Niko: I gotta see what uhh … putting a beard on it- Dean: Oh my god! Nick: This is what the app comes stock with.
Wren: Wooooaaaaaaaahhhh. Niko: Let’s put some hair styles on this. Give him little bangs. [Nick, Nico & Wren laughing] Nick: He looks like a member of BTS. Some k-pop gold, dude. Can we add hair, then screen-cap that then add more stuff and see if we can create a full face from that drawing? Niko: Hairstyles. Let’s give him some bangs. Apply. Let’s give it some glasses. Apply. Nick: Okay, what about a nice little beard? Niko: Nice little beard. Let’s give the [unintelligible] a beard. Apply. [Niko laugh] Nick: What the… Niko: Let’s make it old. There we go. Apply.
Wren: Wooow. Niko: And then I’m going to reimport it.
Wren: All from that? Niko: Okay. So little by little, it’s actually shading it. Like it’s actually filling in like skin colors. Wren: It created those cheekbones from scratch. Nick: That’s pretty amazing. [Laughter] Nick: From that, to that. Wren: That actually looks like, you know those little artist renderings after someone robs a bank? Nick: Pretty amazing that we can take a whiteboard sketch that you just drew and actually make human facial features from it. Niko: So this is definitely doing some sorta like pattern recognition. You can kinda see, kinda see the patterns being repeated. So if you’re like me, you’re probably wondering how this FaceApp program works. Now if you’ve been following this channel for the past few weeks at all You’ve probably seen the deep fakes that we’ve been working on. Deep fakes are basically a type of machine learning Where you are passing an image into the computer It’s running that through its AI and learning on that image and then giving you an AI Generated result that has been informed by the things that it learned on. You also may have seen style transfer. Style transfer is really cool. Style transfer basically determines which parts of your image are just textures and styles and then what do you need to Be able to understand what that image is. Style transfer will basically break those two things apart, preserve the structure of your final image But apply the style from a different image onto that. So you can make every picture look like Van Gogh’s starry night. That’s basically what you’re getting with FaceApp. [Niko laughter] It’s really weird. But what I saw was just wrinkles everywhere. But it’s like there’s a pattern to it. The way this program is working is basically a mix of style transfer with some other structural cues. Kind of like how a deep fake works. If I want to make FaceApp Basically, what I’d do, is I’d first program in a facial landmarks recognition system. The eyes are here. The nose is here. The mouth is here. And then I would train the computer on Here’s what various sets of wrinkles look like. Next to the eyes. Next to the nose. Next to the mouth. And then the computer can Be fed a new face with new landmarks And then go, “Okay, there should be wrinkles here, wrinkles here, wrinkles there. A mustache here. A beard there.” So I did a little experiment to see if I could kind of replicate this effect at very base level. So I gave it this picture of a wrinkly old woman. So I took the style from this picture. Now it’s just a picture. And I took a picture of myself and I applied that picture of the wrinkly woman as the style on my face. And the wrinkles are there. The structure is actually there. Now the background has wrinkles too. It’s all just become wrinkles. So if you were to then inform this with some facial landmarks, you kind of end up with what FaceApp is doing. It’s doing a style transfer of the wrinkles it’s learned from looking at hundreds of images old people. And it’s doing a style transfer onto your skin. When you can combine a style transfer with recognizable landmarks You can end up creating something like FaceApp. You know, I’m almost wondering if there’s like a little bit of image generation. I think it’s called pix2pix. Let’s you, basically as input For example, this purse. And it kind of takes all the pieces of purses and puts them together. So that might actually be what’s going on. I guess I can just literally … I’m gonna draw you Dean. Oh wait, this is for cats. I have no idea what’s gonna happen. I think it’s going to turn you into one of those characters from the new movie, Cats. ♫ Here we go, haha. ♫ [Sam’s Death Squeal] Either way this overall technology is definitely what’s being utilized. The crazy thing here is we are generating imagery in a different way than we used to do it. So when it came to CGI computer-generated imagery. You’d sit there and you’d build the 3D model and trying to make it look real but now we’re getting to the point where [Laughter] It’s horrifying. Okay, I can draw cats I guess. There are thousands if not, millions of pictures of cars, muzzle flashes, phones. Whatever it is that you want. It all exists online. So we can kind of just be like “Hey computer. You’ve seen a million pictures of what the phone looks like. Can you just draw me a new phone?” And it does. You can just tell a computer to piece it together from its knowledge of what things look like. Computer just does it for you and that’s just gonna get better. So yeah, I’m super excited for AI I think there’s so many cool things that are gonna happen. We’re gonna hit a point here, often referred to as a singularity [Laughter] Aw, that poor creature. [Disfigured cat voice] “Help me!” Anyways, you get the idea, it’s AI whatever. It’s crazy. Things are going to change very fast. Clint: Yo, Niko, Niko, Niko. Real quick. I just want to do one thing for my boy. My idol. My hero. (Yeah?) Bruce Lee. Alright, let’s see. Let’s see what our boy Bruce Lee looks like, if he were alive today. Nothing happened. Nothing happened! It doesn’t … it doesn’t work on Bruce Lee. He’s perfectly frozen in time. [Bruce Lee martial arts smacking sounds]


100 thoughts on “We Break FaceApp to See How It Works”

Leave a Reply

Your email address will not be published. Required fields are marked *