the thumbnail of this video is 100% AI generated this is me looking pretty Suave at Paris fashion week this is from another ramp walk from the same event this is me in front of the Eiffel Tower it's a selfie this is me in front of a fighter jet and this is me at a fancy dinner and none of these images were actually shot every single one of them were AI generated and in this video I'm going to teach you exactly how to do this how to train a model on your own face in order to
generate your LinkedIn head shots your Instagram profile profile pictures etc etc but before we do that hit the Subscribe button my name is shev and you're watching the AI Labs by 100x [Music] engineers in order to train a model to recognize your face what you essentially need to do is to create something called as a Laura Laura stands for low rank adaptation it's basically a very fancy way of saying something very simple picture this you have an AI model that is really good at generating images for you but it doesn't really recognize your face so
you add a data set of your face into this existing model and now this model can finally recognize who you are and the reason why luras are a powerful method to do this is because you don't have to train the entire model from scratch again you can just train this new data set that you're adding into the model so it saves a lot of time a lot of costs and a lot of energy it's like if I want a little more salt in my food I'm not going to start the entire cooking process from scratch
I'm just going to take a salt shaker and add some salt to it not a very ACC accurate analogy but paints a decent picture in order to train aura you need two things you need the data set and you need the trainer let's start with the first one a data set is basically a collection of pictures of yourself so I have so many pictures of myself that I have just transferred from my phone now ideally all these pictures need to be in different lighting scenarios and different angles so that the model is able to identify
your face from different angles in different lighting conditions as well I have about 43 pictures but honestly about 20 to 25 pictures will do all of these pictures were taken from my iPhone just make sure they are high resolution and high quality so I have all my pictures in this folder called boson which is a nickname that I use and what I'm going to do is I'm going to compress this into a zip folder now once you're done with this you're done with the first step you have your data set ready now onto the second
component of this workflow the training model itself in order to create a Laura of your face we are going to train it on a flux Dev model flux is an image diffusion model launched by this company called Black Forest labs and it is extremely good at getting realistic faces it's very similar to your mid Journey stable diffusion etc etc and the best part it's open source I've put a link to the flux Dev Laura trainer in the description so you can just open and get started with that now the flux Dev Laura trainer is hosted
on a website called replicate replicate is a website where anyone can basically deploy their AI models for others to use so if you don't have an account just create an account it's free to create an account sign up and you will see the screen that I'm seeing in front of me right now all right now just follow along carefully I will first click the drop- down menu on destination and click on create model and I'll give my model a name I want to call my model B on and I'll just make it private because I
don't want anyone to use it after I'm done with that I will upload my input images so I will click over here and I will upload the zip folder that I just created with my data set now that's done what I'm going to do is I'm going to input a trigger word okay explanation time the trigger word is the particular word that you use in order to recall this data set you're training into the model an example of that would be let's say that the Lura training is done and the trigger word that I used
is boson now every time I want to prompt something let's say I want to prompt something like a photo of boson in a suit now every time I type boson into the prompt or the trigger word into the prompt the model understands that it is supposed to recall this particular data set that I trained as Aura so it is kind of like an identification marker where you're sort of calling back that data set and you're kind of telling the model that hey I want this particular data set this particular Lura that I had just trained
onto this model so if I basically took a bunch of images of my watch and created a data set out of it and let's say my trigger word was iatch every single time I type iatch it basically recalls the training images and generates from that after that we go directly into steps what I'm going to do over here is change the steps to 2,000 which I feel is a pretty ideal spot in order to train these models you don't need to change the learning rate you don't need to change the Lura Rank and now comes
the final two Fields you need to enter in order to start training the HF repo ID and the HF token secret HF stands for hugging face hugging face is a website that is basically like the GitHub for AI you get free models you get free data sets you get free workflows it's a thriving community so what you need to enter over here is your hugging face repo ID and your hugging face token here's how you do that you first go to hugging face. if you don't have an account create one there are two steps you
need to follow in hugging face you look at the top right corner you click on your profile and click on new model all right and just name your model I'm going to name my model Bon and I'm going to make this model private and you should see something like this now I'll just go ahead and click the copy button over here I'll go back to the training window and I will just paste my repo ID over here now now to get the token or the secret you click on the profile button at the top right
corner you click on settings and you go to access tokens now here you need to create a new token you can name it whatever you like I'll call it 100x new and granted permission to everything basically and then you click on create token now this token is visible to you only once so you need to copy copy it because after this you will never be able to see this token ever again so if you did this step wrong it's better to create a new token so I'm going to copy this I'm going to click on
done and I will just paste the secret token over here and that's about it I just need to click on create training and it will actually start training my Laura now if this is the first time you're actually using replicate it is going to prompt you to add your billing details now the cost to train your own Laura or generate images on replicate is very affordable it costs to $4 in order to train your own Laura and for every image you generate it's just a couple of cents so in Indian rupees it's roughly 200 to
250 rupees in order to train your Laura and it probably costs a rupe or two in order to generate an image quite cheap now once your model is training you should start seeing something like this on the log and this process will last for a while about 15 to 30 minutes depending on the size of your data set file now after the training is done you should see a message that says training has been completed all right so we're done actually training aara on your face now how do you actually use this model well I've
actually posted another Link in the description below the training Link in order to actually use the model this is what we typically call as inferencing this is basically what you should see on the screen and I'll tell you how to set it up scroll all the way to the bottom and stop at the place where you see HF laora and all you need to do over here is type your hugging phase repo ID so I'm going to I'm going to go back to hugging face I'm going to go to my profile and I'm going to
select the model and now that my model has been trained I should get all the model weights and the model files over here so I can not only use this on replicate I can use this pretty much anywhere I want to I can click on files and versions in order to see what are the files that has been trained and usually luras come as safe tensor files now you don't need to download this file all you need to do is copy this link just like like you did go to the inferencing model paste it and
that's about it you're done now let's prompt and check it out a photo of boson at Paris fashion week and while this image is loading you can do a bunch of other things on this model as well you can change the aspect ratio you can change the number of outputs you can generate up to four outputs at the same time and you can increase the strength of the prompt or increase the strength of the Laura but whatever default values are already there are pretty ideal for most use cases so you would not need to really
tweak them all right that's me looking like a fashionista at Paris Fashion Week pretty much looks exactly like me now let's try a LinkedIn headshot a photo of Bon wearing a suit and a tie and there we have it I think I look pretty sharp in the Spectra let me know what you you think in the comments now generating LinkedIn headshot is one use case of this model but it doesn't just end there you can use it for your social media posts you can generate marketing campaigns one of the rather interesting things that I did
was basically gather a bunch of amul ads and train a Laura on the amul girl and using this Laura I generated a bunch of images that were pretty similar to amul campaigns this was just basically done as an experiment now one thing that I'd like to mention is that while training faces you can train your own face but before you are using it on someone else's face please take their consent if you do train an AI model on someone else without their consent it can actually land you in serious trouble so stay away from doing
that and stay on the right side of morality and ethics that being said I just have one quick thing to add we at 100x Engineers are doing a hackathon very soon in a few weeks so if you are an AI engineer if you are an AI Builder and if you would like to take part in this hackathon we have put the registration Link in the description as well go ahead and register you can know more details about it in the links that I've given below and I hope to see you in the hackathon until next
time don't forget to hit that subscribe button I'll see you in the next one
