The Free & Uncensored Version of MidJourney! (FLUX.1)

So there's a brand new AI image generating tool in town. There's been a lot of claims out there that it is up to par with Mid Journey and in some cases even does a lot of things better than Mid Journey. So in this video I want to explore Flux One from this brand new company called Black Forest Labs. Before I get into testing and experimenting with Flux One, here's a quick overview of some of the important details that you should know about this specific model. The first thing that you should know is that it was actually developed by many of the team members who helped build Stable Diffusion. Among the team you can see our innovations include creating VQGAN and Latent Diffusion, the Stable Diffusion models for image and AI video generation, like Stable Diffusion xl, Stable Video Diffusion and Rectified Flow Transformers. So the team behind this is a team of rock stars with a lot of experience in creating AI image generation and AI video models. The other important piece of information that you should know is that there are three models on Flux and each one is a little bit more powerful but also a little bit more expensive to use than the last. So there is Flux1snell, which is the fastest model and it's designed for local development and personal use. Most likely, if you're going to run it on a home computer, this is the model that you're going to use. This model is also openly available under the Apache 2.0 license, so it is open source. Any tools that you create that actually use Flux 1 Snell underneath can be sold. Any images that were generated with this tool can also be used non commercially or commercially however you want. Then you have the next model up, Flux One Dev. This is their middle of the line model. It's slightly more efficient than their top of the line model and slightly more prompt, adherent and performant than their Flux1snell model. Now this one can be used for non commercial applications. So you can't create a tool that's using this and then sell access to that tool. And then you've got their top of the line model, Flex one Pro, which is their best model. It offers their state of the art performance and is really designed for enterprise solutions. Now if you want to use one of these Flux models, there's actually a handful of websites that have already integrated it and you can use it completely for free. Right now probably the simplest way to use it is to head on over to Black Forest Labs over on Hugging Face and you can actually use their Snell model, their lowest end model and their dev model within Hugging Face. Spaces. So if I click into this space here, I get a real basic image generator for Flux1snell, I jump back here. You can see we've also got Flux 1 dev model. Both of them are just pretty bare bones. You enter your prompt, we can see where it generates the image. And we have a few advanced settings like a random random seed, width, height and number of inference steps. If you do want to use the Pro model, there's actually a platform called Glyph, which is this really cool AI workflow builder. And you can build your own Flux workflows in here and it will let you generate for free even using the Pro model. So if I click on build up here, I can start building with glyph blocks. We want to start with a text input here. What is your prompt? If we want, we can even run that prompt through an LLM like Claude or chat GPT to improve the prompt. So let's retitle this from input 1 to basic prompt. Let's go ahead and run it through a text generator. Let's tell it to take the following image prompt and improve it so that it generates colorful high contrast images. And then I'll put quotation marks. We'll put our basic prompt in there. Close the quotation marks. Under advanced controls, let's go ahead and set the model to Claude 3.5 sonnet. We'll rename this one from text one to optimized prompt. And now we can add another block to generate an image. So let's go ahead and create our image generator here. And we'll go ahead and have it use our optimized prompt here under image generation model. You can see we have the option for Flux Pro and Flux Snell. So let's go ahead and use Flux Pro. We've got nothing to lose right now because Glyph is letting you actually use this model for free. Let's go ahead and do a landscape 16 by 9 image here. Now we can go ahead and close all of these down. And now it asks what is your prompt? And what it will do is take the basic prompt, run it through Claude 3.5 Sonnet, improve it for contrast and colorfulness, and then output an image for us. And let's go ahead and just do a wolf howling at the moon and run this glyph. You can see it's looking at my basic prompt and then it's optimizing that prompt and now it's generating an image based on that optimized prompt. And here's the image that it generated. Pretty dang solid. Honestly, I can actually open this step over on this left sidebar and see what the optimized prompt was. You can see it is a majestic wolf with illustrious silver and charcoal fur standing atop a rugged cliff, howling at an enormous luminous flower. Full moon. The night sky is a vibrant mix of deep indigo and swirling purple clouds. Etc etc and you can see it got all of that into this. Now I personally haven't spent a ton of time playing around with Flux myself, but a good buddy of mine named Miguel who goes by Angry Penguin PNG over on Twitter, he has spent a lot of time with it. And if you're not following Angry Penguin, I highly recommend you follow him on Twitter. He is at the cutting edge of everything that's going on with AI. A lot of the information that I get before a lot of people get it I get from AP over here on X. So make sure you're following him if you're not. But I called him up because I wanted to get some details from him about what Flux is good at, what Flux isn't great at, and what other details we should actually know when going into use this tool. So let's start with what Flux isn't great at.

It's not good at illustrations. From what I've seen, like it can be better, but I've seen like when people are doing like the fine tunes of like their sdxl, they can get much nicer things for illustrations.

So let's try an illustration. We'll go with a hand drawn illustration of an angry penguin. And this actually looks like a pretty good image, but it doesn't really look hand drawn in my opinion. So I can see where maybe it's missing some of that detail. Let's try an oil painting of an angry penguin. And once again, it looks pretty good, but it doesn't scream oil painting to me. All right, last one. Let's try a watercolor painting of an angry penguin. Yep, that looks like an angry penguin to me. But it's got some watercolor elements in the background and maybe some splashes. But definitely the penguin itself is not looking like a watercolor to me. Not bad. But if I look at what Midjourney generates with the same prompt, a watercolor painting of an angry penguin, it definitely screams watercolor to me a little bit more. Or if I ask it to generate an oil painting of an angry penguin, definitely looks more oil painting. Like you definitely get those brush strokes like you'd see in an oil painting. I didn't see that as much when using the Flux 1 model. So I went ahead and published my glyph real quick and you can see it actually gave me a different user interface. But now let's talk quickly about what Flux is really good at. Is it designed to do realism really well?

Yeah. I would say that the team did a lot of work on aesthetic training. Right. And then just making sure that it produces really great quality outputs, like right out of the box. I would say it's on par with Mid Journey if you're using the right prompt.

So let's test a photorealistic image. Let's try a photo of a man with a beard on a city sidewalk eating an ice cream. And it's pretty darn good. I would say our dude might be sniffing the ice cream more than actually eating it, but it's pretty realistic looking. Let's go ahead and try another prompt. And let's do a woman taking a selfie on a tropical island with the beach in the background. And let's see what we get. Now. That's not bad at all. I can tell it's AI. It doesn't look ultra realistic to me. But that could also be the prompt optimization that I put in. Let's go ahead and build a even simpler workflow where we just enter our prompt here and then just generate the image. So we'll just put prompt and we'll have it just use our input to generate the image here. And we'll use Flux Pro. And now we have a very, very basic image generator. I'll go ahead and publish it so we get that other user interface. And now it's not going to optimize my prompts. This is going to come straight out of the image generator. Let's do a portrait of a woman in a corporate setting about to give a meeting. And we can see it generated a pretty realistic image. Now when I use this same portrait of a woman in a corporate setting about to give a meeting, I feel like Mid Journey still has it beat by a little bit on realism here. These images look much more realistic, much more like something you'd actually find on a stock photo website. But they also kind of look the same. There's nothing that says about to give a meeting in any of these images. Where the glyph version. She's clearly about to give a meeting.

What Flux is really great at is doing anything that has to do with text. So you can make like logos with it. They come out like insane. You can create like Snapchat selfies that look real. I was doing like an example where it was like George Washington crossing the Delaware and as a selfie and it came out like super insane. Just like little things like that. I think it's going to be really great for making memes. The text alone is something that we did not have with SD3. We really didn't have it with SDXL. Dall E3 kind of does text pretty well, but Lux is definitely a step above.

So let's do a polar bear holding a sign that says Mr. E. Flo and it pretty much nailed it on the first try. I don't see anything wrong with that. Let's try this one. A plane writing the word subscribe to Matt Wolf out of smoke in the sky. Let's see how it does with a little bit longer text. I mean, not horrible. It says subscribe Wolfie. It's almost there, but it definitely sort of lost itself a little bit on the longer text here. All right, so running this prompt one more time. This time it just got the words subscribe, but none of the rest of it. I think the plane looks a lot better on this one.

Yeah, I would say prompt adherence is also a huge one. Aesthetic quality. The fact that it's like not censored. Right. Is like a huge thing as well. For people that are creating locally, it is censored. Like it does have an NSFW filter. So it's answered in the good ways, but not in the ways that you know you want to create like maybe a political figure just for fun or something like that. And you won't get hit with like a shadow ban or something like that. So I think just having that optionality is pretty important as well.

Now this one's also supposedly really good at prompt adherence, an area that midjourney is not great at. And when I say prompt adherence, I mean adding a lot of things into the prompt and having it get all of those elements. So for example, a three headed dragon watching TV while eating nachos and wearing a cowboy boots. If I was to plug this into mid journey, I'd likely get a couple of the elements. I'd probably get like the three headed dragon and maybe a tv, but it would miss the nachos and the cowboy boots. That's one area that Dall E3 is really good at. It captures all of the stuff from your prompt and usually figures out how to get it all in there. So let's see how flux one does with getting as many elements into the prompt as possible. So definitely not bad. I wouldn't say this is a three headed dragon. This looks more to me like two separate dragons. But it got the nachos, it got the tv, it got the cowboy boots. Let's give it one more go around. This time it looks more like a one headed dragon eating spaghetti, watching TV without the cowboy boots. So I mean, it's prompt adherent ish. But I'd probably put it about the same level as midjourney when plugging it into mid journey. Here's what I got. One with three dragons eating nachos, no cowboy boots. Here's one with just a normal dragon wearing cowboy boots, eating chips but no tv. Here's three dragons eating burritos or something. And here's three dragons eating chips. None of them got the three headed dragons with the cowboy boots, watching tv, eating nachos. None of them got them. All right, now compare that to what we get when we plug it into Dolly3. Check this out. Three headed dragon eating nachos, wearing cowboy boots, watching TVs. Even got a bowl that says nachos on it. So Dall? E3, when it comes to prompt adherence, absolutely crushes both Flux 1 and Mid Journey. But Dall E3 kind of still sucks at realism. Now the other thing I want to test is the sort of uncensoredness of Flux. Now you can't generate not safe for work stuff yet. It's open source. People are eventually going to break that. And if you want to generate not safe for work stuff, it's only a matter of time. But right now you can't do that. But theoretically you can generate copyrighted images and use existing ip and I want to kind of test that. Let's do SpongeBob SquarePants High Fiving Super Mario. And I mean, the hands sort of merged together, so that's not great. But we definitely got spongebob and we definitely got Mario. Let's see what happens when we try to get like celebrities into it. Let's do Tom Hanks hugging Kanye West. I'd say that looks more like Kanye west huggin an alternate version of Kanye west.

But that definitely looks like Kanye. Let's go ahead and do Tom Hanks standing next to Kanye. That did a pretty good job. The Kanye looks pretty spot on. Tom Hanks not quite there. But the point being it's not going to disallow you to generate whatever you want. If I want Donald Trump eating a donut, I can do that. If I want Kamala Harris eating tacos, I can do that. If I want an image of Baby Yoda hanging out with Spider Man, I can also do that. If you want some additional tips for how to improve your prompts with this a, you can use an AI method where you give it a basic prompt and then AI automatically improves the prompt for you. Or here's some tips directly from Miguel. Let's say somebody just wants to mess with prompting themselves. Like without using the prompt tuner. Like, what are some of the things you've noticed work? Well, because I remember like early on with Stable Diffusion, if you added things like 4K and HD or octane render or Unreal Engine, or you know, you added some of these little keywords to the end, it would actually have a pretty big impact on the output image. Is there any like little tricks or things you've noticed to actually get better outputs from the images?

Yeah, so it does have a dual encoder. So you could use like the simple prompts and the more complex prompts. But if you really want like something super, like insane, I would say try and be as detailed with the prompt as possible. Who puts amazing stuff on X? They are like always pushing the boundaries of the new and upcoming models. They have like some insane prompts that they're putting in like very descriptive down to the word, like spaghetti monster coming out of a donut and you know, attacking something. And it's just like super descriptive. Super. Well, like I'm looking at one right now that's like the word cat where the shapes of all the letters combine to form a cat using only clever typography. And it's like spelling out the word cat with like a cat in there, right? It's like insane what you can do just with, you know, really using your imagination to get something creative.

Now one thing that's really cool about this Flux model is they've already announced down here that this is also going to be a text to video model. So right now Flex 1 is text to image. But they will serve as the powerful foundation for their upcoming suite of competitive generative text to video systems. So pretty soon, using fluxone, you'll have an open source option to generate the types of things that we've seen from tools like Luma's Dream Machine Runway Gen 3 Sora, which nobody really still has access to. Pretty soon you'll have an open source version to do that with Flux 1. Now, does Flux 1 really kill mid Journey like some of the headlines are saying? Not yet. Mid Journey is still creating superior images. Dall E3 still has much better prompt adherence, but Flux 1 is definitely better than what we're getting out of the existing stable diffusion models like sdxl and stable diffusion 3 and they're close, close to on par with what we're getting out of Mid Journey in terms of realism, text generation, and getting a lot closer to the prompt adherence that we're seeing from Dall E3. So this could, if it continues to evolve, which of course it will be a contender to do all of the things that Mid Journey does well. Dall E3 does well and stable diffusion does well. Dall E3 great at prompt adherence, Mid Journey, great at realism, Stable Diffusion great at pretty much uncensored. Anything you can possibly dream up in your mind, you can do it without any guardrails. Make whatever you want. Flux looks like it could be the best of all three of those worlds soon. Keep in mind that the Flux Snell model is open source, so other developers are going to get their hands on it, they're going to fine tune it, they're going to improve it, they're going to rip off the few guardrails that it has now. People are going to create loras for these and you're going to be able to generate better and better images with it all with an open model that's free to use, that's open source, that you should be able to download on your computer pretty soon and run locally if you wanted to. So we're not far off from being able to just install this on our own computers, run them without being connected to the Internet and generate anything you can imagine. So I'm pretty excited about it. Personally. I'm likely still going to go to Leonardo or midjourney or dall e3 depending on my use case. But I would say give it a month or so and this Flux model and whatever people build on top of it is probably going to be generating much better images than those platforms. So exciting, exciting advancement in the world of AI art. I'm super pumped to play around with it more. If I find even more cool tips and tricks and ways to improve my outputs, I'll make a follow up video to this one. But right now I'm just excited that there's a new open source model to compete with Stable Diffusion that already seems to be better than Stable Diffusion. Again, I'll keep you looped in as I learn more. Thank you so much for tuning in. Thank you once again to Miguel AKA Angry Penguin for helping me out with this video and walking me through what Flux is and isn't good at. If you want to stay looped in with all the latest cool AI tools and news, check out Future Tools IO and join the free newsletter all send you an email to your inbox with just the coolest tools and most important news each week. And if you like videos like this, you want to see more tutorials, news and breakdown of the latest happenings in the AI world. Make sure you like this video and subscribe to this channel. I'll make sure stuff like this keeps on showing up in your YouTube feed. Thank you once again for nerding out with me today. Really appreciate you. See you in the next one. Bye bye.