I learned to make Deepfakes... and the results are terrifying

deep fakes have been around for a while namely in Academia and Hollywood by seamlessly transplanting one person's face to another you can make them say something they never said or do something they never did and even resurrect them from the dead thanks to fast graphics cards with lots of memory becoming reasonably affordable and free open source software the power of deep fix has made its way into the hands of a very dangerous group of people the public bestowed with this gift this leap in technology filtered down from the powers that be to the masses what

would we do with it we made porn lots and lots of porn they even got Dumbledore we also made fake political propaganda and many many memes Dragon will do the draw me like one of the French girls you can see how this can be hilarious but also extremely problematic wearing only this [Music] but how easy is it to create a deep fig is it just a drag and drop scenario or do you need a computer science degree to make a meme let's take a journey into the world of deep fakes I want to find out

how long it will take to create a reasonably convincing deep fake perhaps I could become Elon Musk but his incredibly awkward style of talking king or maybe I could become 1980s Arnold where even his face was muscular all I know is this is going to require high level problem solving but dire visual consequences let's get started the program I'm going to use is called Deep face lab it's free and open source and apparently 95 of defects are created using this so it must have a friendly intuitive user interface with an exhaustive manual right well this

is what it looks like just a bunch of batch files in Windows Explorer with no instruction manual as always we'll start with the first tutorial on YouTube alright I have downloaded deep face lab the program with its um very friendly user interface now it is time to start figuring out how to use this program this is day one hour zero all right this is the first tutorial I could find on YouTube always a good place to start it's 17 minutes 41 seconds long let's see how I get on whilst I'm figuring that out let's talk

about how all this works anyway the short answer is it's complicated but here's a Layman's take on it deep face lab uses machine learning to learn what someone's face looks like you feed the program thousands of pictures of the person you are trying to emulate different lighting different angles different Expressions you're trying to tell the program this is what Jim Carrey's face looks like in every possible situation the program trains over hundreds of thousands or even millions of iterations changing model parameters attempting to learn that phase it then replaces the original with the learn face

matching expression and Lighting note here it is not simply photoshopping in different pictures on top of the original it's actually creating a brand new image for every frame based on the model developed by the learning process and that's pretty cool look there's thousands of photographs of me [Music] so apparently all I do now is leave this round for an hour or two come back and we should have a result so I've actually had this running for quite some time now well it's just disappeared I've stopped the model and now I can hopefully package it up

and see a result the idea here is it's learned elon's face it's learned my face I should replace one with the other one I should have a video result let's open the result [Music] but the life cannot just be about solving one miserable problem after another can't that can't be the only thing they need to be they need to be things that inspire you [Laughter] it's so bad yes oh dear okay [Laughter] um I think there's a bit more to this than uh just watching a YouTube tutorial let's go deeper in an attempt to improve

this monstrosity I thought I'd just leave it running machine learning improves using a iterative process so longer is better from what I've read online even using super fast gpus convincing deep fake models can take hundreds of hours of training so I just left it I left this model running overnight and I've just merged the frames together let's have a look at the result maybe this will be better but the light cannot just be about solving one miserable problem after another it can't that can't be the only thing they need to be they need to be

things that inspire you oh did you see that there his eyes inspire you oh dear so it seems like it's fixed the face like coming out of the head problem but it's so blurish that's what I mean it's better we did this uh but it is not convincing at all another strategy required I thought perhaps Elon Musk is just too different to me his Skin's power than mine his face is a totally different shape maybe if I try the different faces of destination I'd have more luck someone with my Olive complexion who could match my

jaw dropping good looks someone like Johnny Depp seriously how can someone be that handsome I got the model up and running and left overnight all right it's the next day I've left that running for approximately 12 hours I have a good feeling about this one I've ruined the joining that space it looks like some sort of weird [Music] and stop all off there's no audio so garbage I tried a number of different things to improve my deep fix what about some 4K footage with some nice lighting for our first trick we'll be trying to connect

a couple of Thunderbolt 3 devices to our USB or laptop and the reason this matters is because how about shot that's dead straight on so no profiles I mean it it's better [Music] not even close I tried everything I could from YouTube tutorials and guides online and everything I made sucked up I can go slow ahead come on down and jump some of this [ __ ] you're gonna need a bigger boat [Music] I know Kung Fu show me [Music] look how they mess because my boy here's Johnny sometimes it just doesn't work after scouring

forums and reading this huge guide I finally stumbled across something pre-training you see when you start training a model at first it doesn't even know what a face is it's a newborn baby with lots of potential but right now it's as thick as mints its first guess or iteration is a blurry nothing it has to learn how Smiles work where the eyes are supposed to be her faces turn Shadows beards blinking all from scratch that's a lot imagine trying to paint someone's portrait if you've never even seen a human face before however if you can

show the model what human faces look like in general before you ask it to copy one for itself it will have a much better chance of getting things right this is called pre-training instead of feeding the model thousands of photos of just one face you give the model 50 000 different faces to look at fading Expressions lighting skin tones and ages you take the model's hand and say welcome to the world this is what humans look like this process can take a long time the model is still just a baby at this stage experiencing its

world for the first time after almost a week straight of doing this the model is now a confident educated teenager now when you tell it to look at my expression and replace my face with Arnold's it can use everything it learned growing up to learn Arnold's unique features focus on now on just that one source the Adolescent model matures into adulthood and can actually start matching Expressions skin texture teeth eyes Shadows it was using this method of pre-training that I got my first decent deep fake all right finally we are getting somewhere because this is

not too bad look that is reasonable like it looks like he's kind of skimped on his diet and uh not worked out in 10 years this is the first time I've actually felt like I've learned something here with all that said let's keep going with something to work with it was time to hone this skill I spent and I really mean this I was reading the Skype I think I've read it top to bottom 10 times there's a lot to this blending masking Source face manipulation different training methods gradually I learned the quirks of the

deep face lab program and things started to look better well this process is incredibly tedious but apparently it makes a difference foreign [Music] thing I've learned is garbage in garbage out if you feed the program crappy Source images then the result can only be crap while namely an Academia and Hollywood by seamlessly transplanting one person's face to another you can make them say something they never said or do something they never did however lots of stuff still sucked sometimes it just doesn't work I need your clothes your boots and your motorcycle you forgot to say

please all right I am ready to present my magnum opus all right so I had to draw a line in the sand here otherwise I would just keep making these forever initially I thought I would learn how to do this for 10 hours and then present my results but I sailed past 10 hours and had nothing good to show for it then 20 30 40 50 I just kept making these but I had to stop eventually so I drew the line at 100 hours of trying to learn how to make deep fakes so I spread

that over about 30 days and whatever you're away to see is the result of that now I had three PCS and a laptop running pretty much 24 7 during these 30 days so it's likely thousands of hours of compute time required to make these and my results are nowhere near as good as a YouTube channel like Ctrl shift face there's a skill to this it's a blend of science and art and I got okay at it but there are Masters of this craft out there anyway here's my best effort here's a montage of me living

Another Life as a movie star enjoy [Music] this is behind the mask you can look into your eyes as you die peace it can't be gulami from now on it's gonna be nothing but short short skirts around the house uh this is Houston uh say again please Houston we have a problem I'd say like one out of a million so you're telling me there's a chance truth is I am Iron Man we're over there now I want to dance I want to win I want that trophy if you enjoyed this video then you might be

the kind of person who is interested in a career in Tech if so I want to introduce you to this week's sponsor Boolean Boolean is an online tech Academy focused on preparing you for the industry they're full and part-time courses consist of live lessons taught by industry professionals with Hands-On job focused projects ranging from software development to data analytics you don't need any prior coding knowledge Boolean will get you from zero to Industry proficient in six months this is not just a bunch of tutorials this is live lessons focused on bringing you up to an

industry standard and once you've graduated they offer six months of career support they also have a financing option where you can pay for the course once you are earning billion understands six months is a big commitment so they're offering you a free coding week it's a five-day event with four days of live lessons so you can start learning how to code and build web apps from scratch this will let you see how Boolean teaches and trains you and find out if it's the right fit for you before you sign up for the six month course

so if this is something you've been thinking about doing this year and want to move into the profession then boolean's free online event could be the ideal first step it starts January 2023 on the 16th and it's free the link to sign up is in the description below thank you very much to Boolean for supporting the show and thanks to you for watching this video I'll see you next time peace [Music]