Flux AI: Creating Realistic Images with Personalization Techniques

Flux AI enables users to train custom models, allowing personalized image generation with their likeness effortlessly.

### How to Train Your Face into AI Models for Ultra-Realistic Images
* 💻  The text provides a step-by-step guide on how to **train your own face into the Flux AI model** to generate **ultra-realistic images**.
* 🎨  The process involves using the site **replicate.com**, creating a new model with **your face**, and setting up the **training parameters**.
* 💰  While the cost of training the model is typically around **$220**, the text offers a **$10 coupon code** to replicate.com, making the process essentially **free for the first 86 image generations**.
* 🤖  The model is trained using **1,000 steps** and can be accessed through Hugging Face once completed.
* 🖼️  The text also demonstrates how to use custom prompts to **optimize image generation**, including creating a project in Cloud to **streamline this process**.
* 🌐  It highlights the importance of ensuring the **trigger word (e.g., "Mr. Eow") is the first word in the prompt** for optimal results.
* 🎥  The final step involves taking the generated images and potentially **animating them** using tools like **Runway gen 3**.
* 👨‍💻  The author emphasizes the **fun and creative potential** of this process, encouraging viewers to **explore AI tools** and **stay updated on AI news**.

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 6'8 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 to and subscribed 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 bulb. 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. 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 Colab. 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 watched away from your computer while it was happening and it completed the process, 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 themselves based on that video, but the process has gotten so much better. So

much easier and so much dang faster. So, we're going to 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 here, I actually broke down that workflow in this tweet here. I used that same file. A site that I used to create the ultra-realistic images, they have this option to train a flux Laura 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 1,000 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 collab 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 2 hours to train the model using this Google collab. Both options work. The Google collab works great. Using file. A also works great. Cost $5. I was able to generate images like this using that method. But in this video, 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 Laura model with our own face into it, we're going to use the site replicate tocom. 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 A00, 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. 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 down until you find this Luca Taco AI toolkit. You can see this is the AI toolkit for Flux Laura training. This is what we want to use to train our Laura this time around. As of recording this video, it's one, two, three, four 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 eflow dlur. 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 to. Now where it says to, 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 eow with little underscores in between each word, and then each one is just numbered after that. Since all of these are named a photo of Mr E flow, it's going to give the model that extra context that when we type Mr eow, 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 a minimum of 12 images is 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 asked 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 huggingface token, so head on over to huggingface doco, 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 discussion permission, so I'll go ahead and name this Mr. Eow 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 1,000 steps. For learning rate, we'll ahead and leave that as the default. Batch size, default. Resolution, default. Defa Laura linear, we're going to just leave all of this to the default. Now when we get down to the repo ID, you can have it automatically upload the Laura 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 Matt Wolf SLM Eow 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. Eow 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, time 60 seconds, time .14. It's going to cost about $210 to train this Laura. All right, 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 $28.18 to do this training run. So now we've got our Laura 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 Laura 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 add 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 Laura. And if we click into this one, we can actually prompt with the Laura 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 can use whatever you want here. You can set your outputs anywhere from 1 to 4. Let's just generate one to start. Infer 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 slmr eow Laura. And then for the Lura scale number here, I'm going to bring this up to 1. 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 eow. 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 theoretically, if I try to run this one more time, I shouldn't get the same error. This time we can see that it is actually properly running through, and there it is. There's what it generated. That's apparently what I look like with

Mr. Eow, as a wizard in colorful robes, looking straight into the camera. Now I'm going to test a few more prompts here in a second, but I also want to fulfill 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 were 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 cost about $210, $220 to actually train the Laura. And then once you're actually using the model, we can see down here at the bottom.

the model costs approximately 9 cents to run on replicate. So, every time you generate an image using your custom model on replicate, it costs about 9. But lwis is hooking you up with a $10 off credit. So, let's say you got $10. Let's say you spend $225 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 Lis 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 Cloud's new projects feature. If I come to Cloud, 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. Eow. 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. Eow, 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, Cloud 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 wanted 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 clad, 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. Eow 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 Dream Booth 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, 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 clad here, and I'll just do something really simple. I'll just put Mr. Eow 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 at 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. Eow, 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. Eow and Deadpool. When I generated 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. Eow and Deadpool, Mr. Eow 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 uh, 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. Trending your own likeness into the flux images is even more fun, and then turning it all into a video, uh, 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, T to TOS, cool stuff like that. Give this video a thumbs up and maybe consider subscribing to this channel because I like to make tutorials like this that are fun and hopefully entertaining and educational, and I'd 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.

How To Make AI Images Of Yourself (Free)

AI, No-Code, Tech, Futurism - I'm a tech nerd and talk about tech nerd stuff

Matt Wolfe

Using Flux AI for Realistic Image Generation

The author has been using the Flux AI image generation model to create realistic images of themselves in various scenarios, including with Deadpool and as Superman.

The model's effectiveness has improved over time, with the author now offering a video guide on how to use Flux for ultra-realistic images.

The process of training the model with one's face has also become easier and faster, with options available for as low as $5.

Training the Flux Laura model using Google Colab or a file costs around $5.

The process can be done for free by renting GPUs on replicate.

com, with a coupon code available to mitigate costs.

The training setup involves creating an account, selecting a model, zipping and uploading images, and obtaining a Hugging Face token for access.

The process involves granting permissions, creating an access token, and setting up the model with default parameters.

The LORA model is then set to upload to Hugging Face, where a private model is created.

After training for 26 minutes, the model's files are dumped into the Hugging Face account, and the repo ID is copied for generating images.

Generating Images with a Custom Model on Replicate

The author encountered an error while running a model on Replicate due to its private visibility on Hugging Face.

After making the model public, it ran successfully, generating an image of the author as a wizard.

The author shares a $10 coupon code for Replicate credits, allowing users to generate approximately 86 images for free after training the model.

The project focuses on optimizing image prompts for Mr.

Eow, ensuring higher contrast, brilliant colors, and beautiful aesthetics.

Custom instructions specify the camera angle and main focus on Mr.

The AI responds with three optimized prompts for each submission, providing options for the best image generation.

The speaker shares their experience generating images of themselves with Deadpool, noting that using their trigger word as the first word in the prompt yields better results.

They demonstrate how to animate the generated image using Runway gen 3, creating a cool slow-motion video.

The speaker expresses their enjoyment of using AI for image and video creation and encourages viewers to subscribe for more tutorials and AI news.