Flux.1 takes AI Image Customization to a whole new Level. [SUPER EASY GUIDE]

So this week is going to be a massive week for AI image generation. There are updates coming to previous AI image generators. There's new generators coming out this week. And today in this video I'm going to talk about what is quite possibly the most versatile image generation model we've ever seen and how to actually utilize it and get that power out of it. You see, I've done a video on this image generator. It is Flux dot one. We all know it, we all love it, but you can actually modify it. It's open source, so we can train it on our own images. And no, you don't need a fancy computer, no, you don't need to have any technical dev experience. It is very simple to set up. So I'm going to show you that towards the end of the video. But first let me show you how you can utilize some of the more customized versions that people have already made. First stop on today's journey is Flux Laura the Explorer. And you can see it's like this little hugging face of Dora the Explorer. This is Free Image Generator, a free Flux image generator on hugging face that has all these different Lauras you can pick from. Even got some decent advanced settings as well including Lora scale. And this will essentially adjust how intense whichever one of these themes you pick out to the left hand side will be. Oh, and just so we can compare, we're gonna go ahead and pull up Grok 2 Mini here on the side which uses the base UNEDITED version of Flux.1 as its IM generator. So let's start off with something super basic. I'm gonna choose Flux Realism. This is obviously for creating more realistic images. Right. First prompt is going to be stunning photography of a red headed woman wearing an aquatic themed designer fashion outfit. Gonna test and push the model a little bit with that weird aquatic themed outfit and we'll see how this Realism Laura compares to just the base version of Flux. And again, keep in mind folks, the Laura the Explorer website here, this is all free on hugging face. It does take a little bit of time to generate. Not as fast as paid models, but it is free and we'll send the same one to Grok and alright, so here is our generated image with the Laura. Looks pretty good. Looks like fashion photography, you know, maybe some fancy designer fashion outfit that is aquatic themed Aqua for sure. And if we go back to Grok, this is what we got out of just the base model with no Laura. And certainly this is no bad image. I mean it follows the prompt pretty darn closely. But I Think there's a better sense of realism out of the Lora just a little bit. To be fair, the base Flux model is actually pretty great at realism right out of the box. But if you want to go just a step further, this is definitely a great option and it's also free. Now let's push it a little bit further. Let's try the tarot v1. So this is obviously going to make, you know, those famous Tarot card style images. And I have the perfect prompt. First of all, we're gonna go to advanced settings and we're gonna increase the height little bit and reduce the width. Tarot card art that represents Walter White from Breaking Bad. Imagery of Walter White from Breaking Bad cooking blue crystal. Let's give that a shot with tarot v1 and I'm gonna try it again with the base version. So here is the Tarot card imagery that we got out of the Laura. It looks freaking awesome. You can see, you know, bald guy, looks a lot like Walter White. It says Walter White at the bottom here, which is pretty cool. I think this imagery is fantastic. Fantastic, actually. Really impressed by this Tarot card, Laura. Now, if we go back to the base version, we can still see that we got a pretty great Walter White image, but it's not really a Tarot card. I mean, it kind of has Tarot card elements. It says Breaking Bad, it says Walter White, but not really what we were looking for at the end of the day. In fact, I think I'm gonna push this image a little bit further with the Laura text at the bottom reads in quotations, the one who knocks and finished product, the one who KN spell pretty much perfect. And it looks like a real Tarot card. Like, I think you could actually make your own customized Tarot card pack just with this. Laura hooked up the Flux one for free on the website. So mighty impressive, and I think it's pretty awesome. I could definitely sink a lot of hours into just making like funny versions of Tarot cards. Anyways, they've got a bunch of different models. I'm obviously not going to go through every single one today. It is free linked in the description, so you guys can try them out. But they've got like synthetic, an watercolor, Disney regular anime. They've got just a regular animation model, PlayStation 1 style, which I'm pretty interested to see, actually. Let's try the ever famous Lemon prompt. Yes, my logo Screenshot from a PS1 game. Lemon character relaxing on the beach wearing sunglasses and. Okay, it kind of did more of like a bubbly 3D art, but this is still pretty cool. Not too bad. Sometimes you have to be pretty specific about your prompting. The Loras sometimes don't work perfectly every single time. But I think regardless, this still is a pretty good image. And it looks more like a PlayStation 1 game than the base version does, as you can see right here. So, yeah, that's Laura the Explorer. This is the easiest free way you can start experimenting with these different customized themed versions of Flux Dot one. And it's great. So let's move on and talk about personalization. This is where it gets really good. So Pietro here on X shows an example of Flux trained on the graphics from a video game called ZX Spectrum. Never heard of it personally. Apparently it's legendary though, so I might get a few angry comments from you viewers. But as Pietro points out, this type of pixel art is so hard to replicate for any other AI image model that came before it. But Flux is just bred differently. And what it's able to pull off here is really, really impressive. The lines and all of the pixel art are very, very much sharp, which is pretty difficult for image generators. You know, actually making it pixel by pixel art. Not perfect, especially, you know, around some of these grassy areas and some of the little nuanced pieces, but definitely gets the job done at an impressive level compared to previous image generators. And as Pietro points out, this specific type of pixel art is very, very unique. There's some characteristics that it was able to capture that you just don't get from the other image generators that came before Flux, because like I said, Flux is just better. It's born different. And a quick search reveals that it wasn't a video game. It was a computer system. 1982. Look, you can't be mad at me for not knowing what this is. 1982 is a little before my time. At any rate, you can see the display here is pretty low res with only 15 colors, and it would make sense. That's why the pixel art is so specific and it's a specific aesthetic that's very difficult for AI models to replicate. But if it's something that you like and you want to create images with it, Flux has you covered for that level of customizability, which is pretty wild, by the way. I don't know who this egg guy is, but he's freaking me out. Is this a good game? One of you has had to have played this weird egg game, so I don't know why he's in a graveyard. But anyways, this post right here is actually the reason that I started to look into the Lauras and making Lauras using Flux one because I was just shocked at the level of fidelity was able to pull. So if we want to do our own customized images, our own generations in Flux one using our own Lora, how do we go about doing that? Well, there's gonna be a ton of different ways, but I'm gonna show you what I believe to be the easiest at the moment, at least that I've seen. If you guys know of an easier way, let me know in the comments below. This right here is Flux Dev Lora Trainer by Austris here on replicate.com I know it looks a little bit complicated, but trust me, this is simple. Training your own la, creating your own one isn't going to be free unless you set it up on your own very powerful computer at home. This is pretty cheap, it's only a tenth of a tenth of a cent or ish per second, something like that. It will not cost you too much to actually train it and it's a pay as you go system. So to do this you will need a Replicate account that's free and easy to create though. And after you sign in with an account and go to the link that's in the description below, you'll be taken to this page. Of course, the first thing we want to do is actually figure out what, what kind of Laura do we want to make? What kind of images do we want to train this thing on? So for mine I'm going to go with Terraria. I'm going to make a Terraria themed Laura. So obviously it's going to copy that classic pixel art design that Terraria has. Of course you can do anything. It could be your own art that you've actually created, it could be your own photos that you've taken or a specific style of photo, specific games, specific styles. That's the impressive part about Flux is you can take it really far. I'm sure you have some ideas already popping into your head, but for now let's continue with how we actually get our images. So I'm just gonna grab some screenshot gameplay images from Terraria to use and I'm actually gonna go ahead and try to find some images that don't have the HUD. Make sure all of your image files are either JPEGs or I believe PNGs also work and we're gonna start putting them into a folder now. You're gonna want to try to go for a minimum of at least 10 images, but the more the Merrier. So now I've got a decent amount of images in this folder. I think it's around 34. And we have to compress this to a zip file. In Windows, this is pretty easy. We just go down to compress to when we right click on the folder and then we click zip file and it makes a brand new little zip which I can then drag over in here. So now my folder is uploaded in here and we actually have to name the model as well. So we'll just call it Terrarium. Now, I will mention with the input images here, you can also add a capture caption text file, which can help you get better quality out of your results. But if you don't include captions, they use auto captioning. So this is more of an advanced step, but it's pretty simple. It's a text file and for each file name you can then have a caption next to it. I'm just gonna use the auto captioning, but this is also available. Next up we have to do the trigger word string. So this refers to the object style or concept you're training on. To not screw up the prompting, you have to pick something that's not really a real word. For mine, I'm just gonna do te in all capitals. They've also got this auto captioning prefix setting. This is essentially just a prefix prompt. So for example, mine, I want it to be pixel art, game, screenshot, and all the different styles. So I'm gonna put this in and just bake that part of the prompt in and you can also have it at the end as well. And you'll notice that I do have my trigger word directly in here and I recommend doing that, but it is optional. Next up, we have steps. They say anywhere from 500 to 4,000. I'm just gonna set mine to like a nice and I'm not going to touch learning rate because I'm not that advanced. I'm just going to leave that. If you're new to training, you probably don't need to change this. We'll leave the batch size as one and then we can also change this integer rank. I'm going to do 64 because I think mine's decently complicated. It will add more time to train though. They also have the optional hugging face repo id, so if you want to upload this to hugging face and then it could be used in something like Laura the Explorer, that's where you would do that along with the hugging face token. So that's pretty much all we need to do. It's very simple. So then we press the Create training button and it goes and starts to train our Lora. This does take time, of course, so you're going to have to let this one sit and brew. This is not something that's going to happen right away for you, but you are customizing a model. It takes time to train it. And I've probably been sitting here for about three to five minutes. And we just started training. It kind of has to set a few things up before it begins to make the Lora, but you can see it's doing about a step per second or so. So at 1500 steps, looks like it' gonna take around 25 minutes or so to actually create the Lora, which really isn't that bad. So I'm gonna let this thing sit in the oven and cook and I'll be right back. And our training is complete. You can see it saved our Lora to a file and right towards the top of the page we can actually download the weights. And this is if we want to run locally the Lora, which is absolutely a possibility. You could use this Lora for totally free and generate everything on your computer, but you do need a decent PC for that. Or we could run the trained model right here in replicate. And this has all of our settings and everything else. So let's go ahead and actually try to run the model. I'm just gonna try a dead simple prompt dog. Let's click the run button and as you can see for the first time, it actually has to load the Lora in. And there's our image. You can see it's not exactly using the Lora just yet. Let's try DOG in the style of ter. Remember that keyword ter? Click run. And now you can see our generation absolutely in the style of TER or Terraria, very much matching the Lora. Let's try modifying the prompt just a little bit more. I'm going to up the Lora scale just a bit more and let's do four outputs. And okay. With these settings you can see it very much looks like Terraria pixel art. But I think I might have set the Lora scale a little too high. You can see it says sane results between 0 and 1 and I'm at 1.23. We'll knock that back down and try this again. It is going to take a little bit of experimenting to figure out the perfect settings for what you're trying to accomplish with the image generator. Training your own Lora and running it is Definitely a little bit more complicated than just using an image generator online. So it's gonna take a little bit of experimenting with the image generation itself to figure out the right settings for you. And we can see here we got just a little bit closer. Definitely, you know, a dog in the style of Terraria pixel art. So I'm gonna switch it up now. I'm gonna go for something that's a little bit more in the style of the Lora that we created today. 16 bit pixel art in the style. Remember, including that keyword is very important. And it's just an image of the Coliseum. We'll say in Rome. I'm also going to up the guidance scale just a little bit. This setting can help it follow the prompt. Default is 3.5. Let's give it a shot. And there we go. I think we're starting to nail the style that we're going for here. It's really starting to come forth. The Coliseum I think was a pretty good one to use. You can see it's always trying to do the HUD up there. That's why I was trying to not use game images that had the HUD or the heads up display. But you could just crop this and it would be pretty good. And we can also try no HUD in the prompt to see if that works. There is another little Coliseum. This looks like something you could actually build in Terraria. But overall, wow. This one actually came out pretty nice. I think you can see where we are going with this. Wow. Really impressed. This is pretty awesome. All right, let's keep going. And I put no HUD in the prompt. I'm going to increase the number of inference steps just a little bit to try to get sharper pixel art lines. Essentially this setting right here is how long is the image being generated for once it begins, let's try Stonehenge. And click run. And Stonehenge seemed to be a little bit more difficult or is proving to be a little bit more difficult for the Laura. Again, prompting can be everything. So I'm going to say 16 bit pixel art of Stonehenge, the famous monument in the style of Ter. No HUD clearly isn't doing much for us. I will say though, with these images. Look closely. The pixel art is actually really quite good for being Terraria. Again, this is small pixel art. That's pretty impressive. And here we go. Getting a little bit closer yet again. This is starting to kind of look like Stonehenge. But again it looks like someone's Terraria house. And I think that is part of the issue here is again, it's going to really utilize the images that you've trained on. Maybe we should try actually lowering the Lora scale just a little bit. Tweaking one little setting can make a heck of a difference. And you can see by tweaking the settings, we are definitely starting to get there. I think Stonehenge is a little abstract for the AI to put into Terraria format. I'm gonna try one more time, but this is looking pretty decent. So I vastly changed just a few settings here, lowered the Lora scale way down and upped the guidance scale a little bit, and we got some pretty different results. Weird little cartoon bear down there, but still pretty cool. I think this one is pretty awesome. It's really taking some creative liberty with the overall 16 bit style and the Terraria Laura. You see, I think this one takes a little bit more form of that traditional Terraria, but either way, really, really cool results. Let's try the moon landing. Oh, wow. We got some really cool, unique results. I think with a more specific prompt, we wouldn't have a moon when we're on the moon. But either way, really, really cool, creative, liberal use of that terraria art style. Pixel art. I think it's super fascinating. These images are starting to come out really nice. I am a big fan. I think we kind of have started to get these settings a little bit more dialed now. All right, now I want it to create a Terraria McDonald's. And here we end up with our final results. And I gotta say, they came out pretty great. We definitely have, like, pixel art McDonald's in the style of Terraria. I think the model has done a pretty great job here. Super impressive. Overall. You can see how much fun and how powerful these Lauras for Flux one can be and also how easy and cheap they are to make. Thank you very much for watching this video. Let me know if this helped you out. And also let me know if you want to see a video on how to actually apply the Loras that you make to a local model and run everything for free locally on your own machine. That's definitely something that's also possible and that also might give you more controllability as well. I'll see you guys in the next video. Goodbye.