How To Make AI Images Of Yourself (Free)

Okay, so this is super cool. I figured out a way to use the new Flux AI image generation model and actually train myself into the model so that I can get pictures like this of me and Deadpool walking away from an explosion. You know, ignore the fact that it makes me look like I'm 6 foot 8 tall. It's got the face pretty dialed in. I also managed to make this where it got both my face and text all in the same image. I mean, it missed the word and subscribe to Matt Wolf, but the vibe is there. Here's another one where I'm holding up a sign that says subscribe and it looks pretty solid. Or here's me as Superman above a city. And another me as Superman above a city. And another option for me as Superman above a city. Here's another one that I made of me in outer space as an astronaut. And here's me as whatever this was. I don't remember what I was thinking. I made it really late at night. And then here's another of me in like some sort of workshop with a light. Bul I was going with some really weird prompts. I don't know what I was thinking. Anyway, in my opinion, Flux is one of the best models at generating realism that exists right now. It's pretty on par with what we get out of Mid Journey. In fact, I made an entire video here called Easy Guide to Ultra Realistic AI Images with Flux. And if you want to learn how to make super realistic images with Flux, you should definitely check that out. In that video, I used a tool called Foul AI. They have a model called Flux Realism Laura, where you can enter a prompt and it tries to give you as realistic of an image as possible. Again, check out that other video. I break down the whole process. In this video, I want to show you how to get yourself into the images. Now, back about a year ago, I made this video called Inject yourself into the AI and make any image with your face 100% free method. Back when I made that video, the model we were training into was stable diffusion 1.4. Since then, we've gotten stable diffusion 2, we've gotten stable diffusion XL, we've gotten stable diffusion 3, and now we've gotten Flux. So quite a bit of time has passed, but the models have gotten a lot better. So now we should be able to train our faces into the newer, better, more performant models. Also, back when I made that video, the process to do that took a good two plus hours. You had to train it on Dream Booth inside of a Google Collab, you had to keep the browser open and kind of scroll the website every once in a while to make sure the website didn't time out. If you walked away from your computer while it was happening and it completed the process and it might still time out after completing the process and the weights that you created would disappear. It was kind of a mess. I mean, it worked. A lot of people managed to generate AI characters of themself based on that video, but the process has gotten so much better, so much easier and so much dang faster. So we're gonna talk about that in this video. Now, when I made this image here and then eventually turned it into a video, I actually didn't use the method that I'm going to teach. If you want to see the method that I used, I actually broke down that workflow in this tweet here. I use that same file AI site that I used to create the ultra realistic images. They have this option to train a Flux Lora here and it cost pretty much $5 exactly to train the model. It's a half a cent per step, but the recommended amount of steps to train is a thousand steps. So it comes out to pretty much $5 exactly to train on this website. There's also a Google Collab available where you can run all the steps in the Google Lab to train your own model. When I actually went through this specific process, I ran it on an A100 GPU up here, which does require a pro plan at $10 a month in order to do. And it took about two hours to train the model using this Google Collab. Both options work. The Google Collab works great. Using file AI also works great, costs $5. I was able to generate images like this using that method. But in this video I. I'm going to show you a way that you can do it for free. Now, it's not normally free, but I'm going to hook you up in this video. So stick around to the end and I'll make sure that you understand how to actually do this process with no out of pocket cost to you. So in order to train our Flux Lora model with our own face into it, we're going to use the site replicate.com now this isn't a free to use site. You're basically renting GPUs from them to run the processing. And if you're going to use an Nvidia A100, which is what we'll use to train this Laura, it costs about a tenth of a penny per second. Or roughly $5 an hour. However, I'm going to hook you up with a coupon code that should mitigate any cost of actually training this Laura. So when you're on replicate.com make sure you create your own account. And once you have an account created, head on over to the Explore page and search out Luca Taco. L U C A T A C O. You could pretty much click on any of the models that show up. I'm just trying to get to their username. I'm going to go ahead and click on their username here, Luca Taco, and then scroll. You find this Luca taco/AI toolkit. You can see this is the AI toolkit for Flux Lora training. This is what we want to use to train our Lora this time around. As of recording this video, it's 1, 2, 3, 4 down the page here. So I'll go ahead and click into this and we get a screen that looks like what you see right now. I'm going to click on the tab up at the top that says train. And this is where we're going to actually set everything up for the training. And it's real quick, real easy. You don't really need to change a lot of the default settings. Now, under Destination, go ahead and select Create a new model and then enter the name of the model that you're going to create. For this one, I'm going to go ahead and just call it Mr. E Flow dash Laura. And then it asks for an images file. Now, this needs to be a zip file of the images that you want to train it on. And you can see here the instructions say file names must be their captions. So, for example, a photo of Tok. Now, where it says tok, this is going to be your trigger word that you're going to use to invoke your likeness into the image. So here's a bunch of headshots that I have on my computer. And you can see I renamed them all. A photo of Mr. Eflow with little underscores in between each word. And then each one is just numbered after that. Since all of these are named as a photo of Mr. Eflow, it's going to give the model that extra context that when we type Mr. Eflow, it brings in an image with my face on it. Now, I have 20 different images here. I think the minimum is about 12. Yeah. So minimum, 12 images required. I have 20. So I'm just going to go ahead and use all 20. Once you have all 20 of your images, go ahead and zip them up. I already created a zip file here, but this zip file you can see just contains the 20 images that are right here. So I will take my zip file, drop it into where it asks for our images file under t model name here. We're just going to leave that blank. We'll let it use the default. Now we need a Hugging Face token, so head on over to Hugging Face Co create a free account if you haven't already. Go up to your profile. Click on Settings over on the left, click on Access tokens up at the top. Go ahead and create a new token. I'm going to set up one of these fine grain tokens and these are the permissions I'm going to give it. I'm probably giving it more permissions than it actually needs, but to be honest, I'm not 100% sure what permissions it needs. So let's just go ahead and give it that. I can't imagine it needs billing or disabled discussion permission. So I'll go ahead and name this Mr. E Flow Laura Replicate and we'll go ahead and create the token. Copy our access token here, jump back over to our replicate page and then paste in our Hugging Face token right here. Now under number of steps we want 1000 steps for learning rate. We'll go ahead and leave that as the default batch size, Default resolution, Default Laura Linear we're going to leave all of this to the default. Now when we get down to the repo id, you can have it automatically upload the Lora to Hugging Face so you can access it from there. So I'm going to go ahead and do that. You can see the formatting should be your hugging face username followed by whatever you want to call it. So I'm just going to call it mat wolf/ Mr. Eflow Laura. Now once you've entered a repo ID here, you're going to want to go ahead and create that repo over on Hugging Face. So if I jump back to my Hugging Face account here and I come and hover over my profile pic, I can select new model. And then under model name we're going to use this same name here. Everything after the Matt Wolf slash here, this Mr. Eflow Laura, go ahead and paste that into the model name here. I'm going to set this as a private model and then we'll go ahead and create the model and then we can go ahead and click Create Training. And we can see it's going to start running the training Now. Last time I did this training Process, it took about 24 minutes. And if you remember, it costs about a tenth of a penny per second. So let's just say 25 minutes times 60 seconds times.0014. It's gonna cost about $2.10 to train this Lora. Alright, so the model has completed training here. You can see that it took 26 minutes to train. We do the math on that real quick. That would come out to about $2.18. And to do this training run. So now we've got our Lora created. If we jump over to our hugging face model that we made earlier, go ahead and refresh this. Now if we click on files and versions, you can actually see that it dumped the files, the Lora files that it created into this hugging face model for us here. I'm going to go ahead and copy this repo ID here. We're going to need it in a second when it comes time to actually generate the images. So now it's time to actually run the model and generate some images with our likeness. See how well it did. So once again, I'm going to go to Luca Taco's account here, click back into his profile and at the time of this recording, the very top left option here says Flux Dev Lora. And if we click into this one, we can actually prompt with the Lora we just created. So we'll come back to our text prompt up here in a second. Let's set an aspect ratio of 16. 9. You could use whatever you want here. You can set your outputs anywhere from 1 to 4. Let's just generate one to start inference steps. We'll leave it at the default 28 guidance scale. We'll leave this at the default as well, just to see how it does. You can choose how you want your image to be outputted as a webp, a JPEG or a png. Let's go ahead and do jpeg. And then down here under HF Laura, this is your hugging face repo that we just copied earlier. So we'll just go ahead and delete that and we'll paste in the proper repo here. In my case, it's Matt Wolf, slash Mr. Eflow, Laura. And then for the Laura squ scale number here, I'm going to bring this up to one. And now we can jump all the way to the top and pick a prompt here. I'm just going to start with something really, really simple. Let's do Mr. Eflow. That's my trigger word that I created earlier as a wizard in colorful robes looking Straight into the camera. Let's just go ahead and start with a very simple prompt and let's see if it actually figured out what I look like. We'll click run here. And I actually got this error here. I was going to edit this out, but I figured I would leave this in just to show you what I did wrong, just in case you run into this error as well. When I created my model over here on Hugging Face, I set the model as a private model. So if I go over here to Settings, you can see the model. Visibility is private, which is causing issues with Replicate actually being able to see the model. So if I go ahead and click make public now, this model is public for anyone to use. I can see it confirmed here. The model is currently public now. You're ready. Likely if I try to run this one more time, I shouldn't get the same error. This time we can see that is actually properly running through. And there it is. There's what it generated. That's apparently what I look like with Mr. Eflow as a wizard in colorful robes looking straight into the camera. Now, I'm going to test a few more props here in a second, but I also want to fulfill on my promise to show you how you can do this for free right now. So Louis C. Here, AKA Luca Taco, who actually put this stuff up on Replicate for you to use. I actually was talking to him behind the scenes in some Twitter messages and told him I was about to make a video showing how to train yourself into Laura's Replicate did not sponsor this video. There was no deals made or anything like that. I just told him, hey, I'm going to make this video showing how to do this with Replicate. And he said, if you do, here's a $10 coupon code that people who watch your video can use to get $10 in credits to Replicate. So again, if we're doing the math, it costs about 10, $2.10 $2.20 to actually train the Lora. And then once you're actually using the model, we can see down here at the bottom, the model costs approximately $0.09 to run on Replicate. So every time you generate an image using your custom model on Replicate, it costs about 9 cents. But Lewis is hooking you up with a $10 off credit. So let's say you got $10. Let's say you spend $2.25 on actually training the model. You now have $7.75 divided by 9 cents per generation. You should be able to generate about 86 images of yourself in whatever scenario you can imagine with those free credits. Now, in order to get those free credits, I put a link to Replicate in the description. It's not an affiliate link. I don't get a commission. I didn't make a penny off of talking about this. But there is a special link to use Replicate in the description. When you click that link, you should see a page that looks like this. Welcome to Replicate. You've been given $10 in credit to run and fine tune models. Accept the credit, it'll put $10 in your account. You're good to go. You don't even need to put in credit card information. You can start testing and playing with the model right away. So thank you so much to Replicate, and to Louis Luca Taco here for hooking everybody up. So now that you know how you can use it for free, let's go generate some more stuff and see what kind of cool, crazy creations we can get out of this. So one thing I really like to do, and this is just sort of a bonus tip, is I like to use Claude to help me sort of optimize the prompts and make cooler prompts that are going to get cooler outputs. A really easy way to do this is to use Claude's new Projects feature. If I come to Claude, go to Projects on the left, create a new project, and let's just call it Flux Image Prompt Optimizer. You can name it whatever you want, but that's just a very literal name for what I'm going to be using it for. We can create the project. And the only reason I'm creating this project is I want to set custom instructions for whenever I use this project. And then every time I use this project in the future, it's going to use the same exact custom instructions. It's going to know what to do. All right, so I went ahead and I wrote up a set of custom instructions here. I'll just read it real quick. You are an AI image prompt optimizer. Your role is to take the prompts that I give you and optimize them so that the image generated is higher contrast, has more brilliant colors, and has beautiful aesthetics. The subject of the prompt will always be Mr. Eflow. This is the trigger word to use my face within the image. The prompt should always mention what camera angle should be generated. We want the subject, Mr. Eflow, to always be the main focus of the image and his face to be seen in the image. And then I went on to add, whenever an image prompt is submitted, respond with three optimized prompts to get a better version of the same idea. Don't give any extra context. Just reply with the optimized prompts. A lot of times Claude will reply with, okay, here is the optimized prompts for you. I hope these work. And then it'll give the prompts, and at the end it'll say, let me know what you think or if you want them to be improved. I don't want any of that extra stuff. I just want it to reply with the new prompts. Also, I thought, let's have it do three instead of one, and we can either try all three or just pick the one that generated the best prompt. Just gives us some options. So I'll go ahead and save these instructions. And now let's take this exact same prompt that I used earlier. I'll just copy this one here, jump over to Claude, paste it in, and let our custom project that we built here do its thing. And just like that, we got three optimized prompts. Close up portrait of Mr. Eflow as a powerful wizard. Piercing gaze directly at the camera, wearing vibrant iridescent robes with intricate blah, blah, blah, blah, blah. All right, let's just copy this first one. Paste it in right here. Run it again. Bada bing, bada boom. Look at that image. That is absolutely awesome. Quite a bit better than anything I've managed to get out of stable diffusion using my custom Dreambooth model. Let's try the second prompt it gave us. Here, plop that in there. And we'll run the second one. Here's what that second prompt gave us. The beard's a little bit longer than I've had it for a while, but looks pretty good. And here's the third image from the third prompt. Got an extra finger on one hand, but it's definitely a little bit more intricate and cooler than what I came up with with my original basic prompt. Now let's come back to Claude here, and I'll just do something really simple. I'll just put Mr. Eflow as a basketball player. Let's see what it generates. So we got three prompts here. Let's go ahead and one at a time, see what they look like. Here's the first image that it generated. Here's the second image that it generated. And here's the third image that it generated. Pretty cool. Here's an image I tried to generate of me walking next to Deadpool, but it kind of put me facing the wrong direction. And then here's another attempt where it actually got me and Deadpool in the image with an explosion in the background. Now, one thing I've noticed, and I don't know if this is how this actually is supposed to work, I don't know if this is common knowledge, but when I put Mr. E Flow, my trigger word, as the first word, it works a lot better than if it's somewhere else within the prompt. So this prompt originally said low angle view of Mr. E, Flo and Deadpool. When I generated it the first time, it literally just had an image of Deadpool. I wasn't in the image when I reworded it, so it said Mr. Eflow and Deadpool. Mr. Eflow being the trigger word. As the very first word, it prompted more of what I was looking for. Again, that could be totally coincidental, but I have noticed with multiple other prompts that if my trigger word isn't the very first word, it messes it up a little bit. So for me, best practice from what I've been doing has been to put it as the very first word to get the best result. And then once you do have a really cool shot, you can take it to the next level by jumping over to Runway Gen 3, clicking on Get Started and tossing your latest creation in here to animate it like I showed in my previous video. I usually just take the same prompt here, copy it, throw that into Runway. I'll set this as the first frame instead of the last frame, and we get this epic video of me and Deadpool, slow motion walking away from an explosion just like they do in the superhero movies. And it looks pretty dang cool. So, yeah, Flux is fun. Training your own likeness into the Flux images is even more fun. And then turning it all into a video, it's kind of awesome. I'm having so much fun playing with this stuff right now. Anyway, I hope that tutorial was helpful for you. If you like stuff like that and you want to know about the latest AI news, tech tools, cool stuff like that, give this video a thumbs up and maybe consider subscribing to this channel. Subscribe, because I like to make tutorials like this that are fun and hopefully entertaining and educational. And I like to keep you in the loop with the AI news. And if that's what you're into, I think you'll like being a subscriber. Anyway, that's all I got for you. Really appreciate you. See you in the next video. Bye Bye.