Opensource, Uncensored, Unbothered. - Flux.1 Image Gen

Open source AI has been on an absolute roll lately. Recently we've seen the release of Llama 3.1, some staggeringly capable LLMs. Two weeks ago we saw the release of Aura Flow, which is an open source AI image generator. And now we have Flux 1, which is another open source image generator. But this one is even better than Aura Flow. The text rendering on here is some of the best, the most capable that I've ever seen on an image generator before. It's really, really quite mind blowing. You see, AI is cool and it's all these gnome creatures or whatever holding the signs. Text is like seriously spot on good. Also really good at complex compositions, stuff that you know, probably isn't even in the database. So this is something that classically Dall E3 has been really good at. But yeah, we've got all these people, you know, swimming in this hot tub that's like a giant teacup or something like that. Anatomical accuracy as well. Those fingers are going to look good, the people are going to look good. And I think you'll you' see here that it is very competitive with Midjourney, is very competitive with Dall E3 on pretty much all of these fronts, which is really, really impressive. And like I said guys, this is open source, so it can be built upon, it can be adjusted and also it is quite uncensored, which is something that we're going to dive into. So there's a few different places that you can use this model. Right now I'm going to link all of them down in the description below. Some of them have limited free access, some of them are completely free access with wait times etc. It is open sour though, so we'll probably see it expand to a lot of different corners of different AI image generators. Anyway, so here's my first prompt I'm showing you. This is Kendrick Llama. So it essentially just produced a Llama. This is kind of an inside running joke on the channel. Anyways, let's actually do an impressive prompt here. Fisheye macro view of a grumpy old goldfish in a fish bowl. 3D speech bubble reads these young guppies and their new AI tech. Right, so a pretty complex sentence. Now for options with this AI. We actually do get quite a lot. Look at this. We have all these different aspect ratios. You can even do custom pretty much any value you want, which is great to see. I'm going to do portrait for this one and then you can do inference steps all the way up to 50 at least on this website. And I've been leaving it pretty low. It seems to like around 20 steps or so. And CFG scale, I think they had it pretty low initially off the bat, but I upped it to like six, six and a half, something like that. They also have sync mode, which is something for APIs. And you can do number of images. Of course. They also have safety tolerance here. I'm going to lower this all the way down to the bottom here. We are not too worried about safety and we will click the run button. Another thing you're going to notice about this image generator Flux is very fast and it actually has two different versions. So it has like a more lightweight, fast version that generates in just a couple of seconds and then it has this heavier pro level version. Oh, well, we did definitely get our goldfish. No text here. That is a little strange. All right, I put the safety back to default and it seems to have worked fine. So maybe don't touch the safety setting. At least on this website. It could screw up your text anyways. You can see it. Got it. Almost perfect over here. It says these, these young guppies in their new AI tech. And I like the speech bubble. It's very close to being perfect. And you can see it's got like this old goldfish. He's kind of sticking out of the fishbowl, which is a little annoying. But this, you know what, it's not a bad generation. And re rolling the generation. We can see it does it again. This is an even better image, I think. You know, we have close up of the goldfish's face. We've got that 3D speech bubble. These young guppies in their new AI tech. I mean it's like a perfect speech bubble, honestly, perfect text right off the bat. You think Dall E3 can get this right right off the bat? It might be able to, but not sure. I'm also going to toss it an ideogram as well, which is classically very, very good at text. You can see all 4times 4ideogram AI. It pretty much got it right. Just a little bit of screwed up ness in the guppies over here. But other than that, these are all really consistent and really good from ideogram AI, but still not as cheap as the brand new Flux. And you can see Dall E3 with our first try here, did not quite get it correctly. You can see the text is a little bit screwed up. So again, let's just reroll the dice another time. Didn't quite get it perfect this time, but it's pret dang close for an AI model that is open source right out of the box. Incredibly coherent. All right, so now I do want to move on to the whole safety thing. I'm gonna put it all the way to its lowest. Let's just try some famous people. Some famous figures seems to have worked quite well in this circumstance. Let's try another celebrity. Seems to do quite a decent job. Very natural looking people. I will say that. Let's try combining a famous person and a famous person property. And I think it actually did a really, really good job here. Anatomically, the hands look perfect, and it's literally just spider man with that. That exact hair. Man. That's crazy. That actually came out way better than I thought. Let's try another celebrity now. All right. This time we kind of just got a normal spider man. I mean, it did a really, really good spider man, but hands are a little messed up, and it's definitely not Willem Dafoe. Let's just try adding that. We can see his face. Ooh. So it doesn't know this celebrity, but still not a bad spider man image overall. Good anatomical hands, good face, and good logo. Let's just try, like, the Coca Cola logo. Oh, gosh. That is kind of insane. We could see all these cans of coke, and it's just a perfect little bottle of Coca Cola with a perfect logo. So it's definitely got lots of copyrighted material in there, which means for most of us, lots of fun. But you don't want to be malicious with it because that can ruin the fun for everybody. Kirby sitting in the oval office in a suit as president. Oh, folks, that is not Kirby, I don't think, interestingly enough, Kirby from Nintendo. Kirby. Oh, there we go. Okay, so it's kind of just Kirby's entire body as the president's head sitting in the oval office. Not bad, though. We can see his nameplate that spells President Kirby. Okay. This is maybe a little bit better still. Really off putting and creepy. The hands are not too bad, though. And obviously not spelling president Kirby perfectly there. Just up the CFG maybe a little bit and rerun it. There we go. Upping the CFG gets us some better results here. We could see Kirby kind of wearing a suit. I mean, that's not too bad, right? President Kirby. There he is. It gets the job done. This model. I mean, it's really, really impressive. Get this. You know, maybe in Dolly 3 inside of ChatGPT. Probably not. Yep. Can't generate it because it's not matching the Content policy Ideogram will probably allow us to generate this though. So let's give it a shot real quick with Ideogram. Does Ideogram do a better job? Not really. I think the president, Kirby looks better, and it's very, very consistent across all of these. But Kirby himself isn't wearing a suit at all. And it's definitely like this really weird, like, off brand version of Kirby. It's like, ye, don't worry, we have Kirby at home. And then this is Kirby at home. Like, why does bro have a nose in two of these? Something ain't right here, man. I gotta say, I like this. Now we'll do Obama posing for a photo with Kirby. And dang, that's pretty good. It kind of looks like a Kirby stuffed animal, but that's really not that bad at all. I'm actually really satisfied with this. This is quite good. Now they're posing, you know, for a nice photo together. Can we add Walter White into this and do this? Dynamic trio ultimate rotation. Okay. It looks like it combined Obama and Walter White together and then put Kirby on someone's head. Let's give this another shot. Maybe we'll go a little bit wider. These prompts that I'm running, I want you guys to understand that they are not easy prompts. These are difficult prompts for the AI to get correctly. Not easy by any means, but hey, there we go. All right, we got Kirby, Obama, and Walter White posing. The Walter White kind of doesn't. It's not perfect, but it's good enough. Like, this is a. This is definitely a good image. Makes me happy to see they're all hanging out together. Can we make this even wider? And Super Mario from Nintendo Mario. And let's add a fifth person in an anthropomorphic lemon character wearing sunglasses. This prompt's getting really complex now. The flux just. I mean, it keeps impressing me. I've been pushing it all day. Now I'll show you some of my older gen generations as well and some other things that you can do with this thing. But it's really, really shockingly good. All right, we got a pretty decent Obama. This looks like a minion, some weird lemon dude, and fake Kirby. Again, this. This is really starting to push the model. We gotta try to get it to produce some good results. See, it's starting to struggle. Now that we have five different characters. Obama keeps looking pretty good. But, you know, we got very, very yellow Mario wearing sunglasses. We got Kirby in the background and this creepy guy. Suggest this prompt. Okay, again, we Got Mario, we got Obama, we got this Lem. It's starting to fall apart. I think the Lemon character might be screwing it up and. Okay, that's close enough. Very good, Obama. Kirby's looking a little whack. I don't know what they've been giving Kirby again. That is definitely Walter White that we have at home. Not Walter White from out and about and not Mario, but Luigi. So you know, some variation in there, but overall the composition of the image is quite good. Anatomically good and very good in terms of getting all those characters in there, even if it's not perfect. Let's again we're going to try this in Ideogram AI, which I consider to be probably the best uncensored model we can get our hands on. That is not open source. I would say does about just as decent of a job. I think the Walter White definitely is looking a little bit worse in all these examples. Again, Kirby's looking pretty knockoff. Obama looks good. Mario looks good in all these. It's close, it's close. But the fact of the matter is folks that Flux is open source source. This model is not the same as Ideogram. You can't really compare them in that same sense if they're about as good in image quality. The open source model in my mind is always going to be cheaper and more accessible and have fine tunes and stuff like that, which is more valuable at the end of the day. Let's try some more boring stuff instead of these crazy images. Logo for a coffee shop that is entirely run by fish gave us like this interesting fish logo. Text on logo reads Coffee for fish by fish. Okay, got pretty close here. I think we'll have to up the cfg. The CFG seems to be the most important setting for your sensitivity to text and overall prompt following. But again, you can really deep fry images with CFG scale if you don't use it correctly. So be careful and watch out for that. Yeah. Coffee 4 Fish by Fish. That's a perfect logo. So yeah guys, this is Flux 1 Pro on Fall AI. You do have to buy credits with Fall AI to be able to use this thing, but I am going to link it down below. You should be able to also use it on the flux.org website here. So let's just run a quick little prompt. We do have advanced settings here too, which is nice. You can see this takes about 20 seconds coming from hugging face I believe. And by the way, this Schnell model, this is the smaller model that generates more quickly. Still delivers in my opinion, really, really great quality. And there we go. Image of a cat. So I think this is entirely free. I don't know if it runs forever, but yeah, this is the Chanel smaller faster model. You can also access the smaller faster model Chanel on this website right here which is a hugging face space. It's the official Black Forest Labs one. So this one in my opinion is a little bit faster than the flux.org use website. You'll notice I click the run button here, it's about a 10 second wait in and we get our image in actually less than 10 seconds here. And you'll see it got the text pretty close to being correct. Even the image and like the detail and the quality and stuff is not bad here on this one I don't seem to have any access to CFG scale. So let's rerun it, see if we can get some perfect text. I have gotten plenty of really, really good quality text. About as good as Dall E3 or Ideogram even I would say from the smaller Schnell model. And again pretty close with this one, pretty close with this one. So the Pro is going to be a little bit better. But this model by no bad and it's still plenty uncensored and still also plenty of ability to generate celebrities as well. And this is an example, as you can see, generating a pretty perfect Morgan Freeman tiny astronaut hatching from an egg on the moon. Oh, that one came out really, really nice. Cat holding a sign that says hello world, he's got three paws. Unfortunately anime illustration of a wiener schnitzel. And of course folks, we do also of course have the ability to do image to video which is really, really intriguing. So you also come out pretty good. You can generate your image initially with Flux and then you can parse it and input it into Gen 3 Alpha or cling AI any of the other generators and get some pretty great video out of them as well. And of course like if I were to prompt this the same way, you know, Obama holding a burger and taking a big bite, Gen 3 would never let me do that. But you know we can circumnavigate that by using Flux to make the image first and then put it directly in here. So that's a little tip for you guys as well, something that I've been doing lately. You can also access the Flux model here inside of replicate.com. this is another API website. So I believe you get some level of free use out of this. But then you have to start paying and you can see again, really, really great examples here. I mean this is the Chanel model, this is the smaller model that generates in less than 10 seconds. And just look at the quality of this thing. It's really, really stunningly good. Now the people behind this Flux model, Black Forest Labs apparently are also working on some AI video generation model to compete with the rest. I will say the video generation is looking pretty good off this little demo, but this is all we have. It's just like this little teaser, state of the art, text to video for all apparently. All right, I'll see it when I can use it, you know what I mean? But like I said, these little demo videos that they have here are quite impressive. So let's talk a little bit about licenses and how you can actually use these open source models. So Flux Pro is going to be their API model. Then they have the Flux dev and this is the actual open source version of Flux 1. It's open weight guidance distilled for non commercial applications unless you contact them first. So it can be open source for commercial purposes. You just need to get in touch with them. You can't just go off and do it. Not a huge fan of that aura. Flow actually does have a better license in that sense, but they still both are open source, which is good for the community regardless. In the smaller Flux1 Schnell model, the very fast model, it's still quite capable. And this is Apache 2.0, so that's really nice. The code, the weights can all be found for this guy and you can use it for whatever purposes you want. So Schnell is fully open source.