Easy Guide To Ultra-Realistic AI Images (With Flux)

Man, these AI generated images have been really good lately. I mean, check out what we're getting. We got this one. I mean, this one is just phenomenal. Check that out. Like, if you were just scrolling on Instagram, you would never know that that was AI generated. We've got this work of art. I mean, fantastic. I cannot find anything wrong with this image at all. No, no, I'm just screwing with you. These are all from stable diffusion 3, which from this point on will probably always be known for images like this and things like this. But truly for reals, AI images have gotten really, really good. I talked about this in a recent news video about how flux came out and people were figuring out how to make Flux even more and more realistic. We got images like this and like this one and this one here, and here's another one. I mean, as you can see, it's getting harder and harder to tell when an image was generated with AI. These are all flu and flux is just absolutely insane at creating super realistic images. I think the fact that they're not like perfectly composed, like they don't look like a professional photographer took them, is sort of what gives them that feeling of, all right, this looks like just a random snapshot that somebody took. It looks real. Like if you were just scrolling Instagram and you saw this without looking super, super closely, you probably wouldn't know that was AI. I mean, look how it's sort of off centered, like someone taking a quick iPhone picture would probably do. Now, these are all really, really good. There are a few exceptions of when it starts to get really kind of wonky. And that's when you try to get more of the body in the shot. Then the proportions start to look a little bit off. But even that often just takes a few RE rolls and you get something that looks decent. Now, I didn't generate any of these, by the way. These are all ones that I found on Reddit. But then people on X started taking this to another level and taking these realistic looking images and, and animating them and making them into realistic looking videos. You see a video like this and you know there's no sound to it, but you just scroll past this. This does not look AI generated to me. Here's another one of an AI generated woman talking on stage and then the same, like ponytail dude over on the right. And once again, this was generated with flux. And then it looks like they used Luma's dream machine to take that image and turn it into a video. Here's another video that I came across that looks like somebody at a TED Talk and their paper actually has the Now I kept on coming across a lot of this stuff and I was having a really hard time getting the images that I generated with Flux to look ultra realistic. I was using this Glyph app workflow builder here because it actually lets me use the Glyph Pro version for free. But when I run a prompt through the Glyph version of Flux, I'd get images like this, which to be honest is actually really, really good, really realistic, but it's got this like almost plastic shininess to the skin that we weren't in some of those other images. Here's another output that tried to generate that same guy and once again it is pretty dang realistic, but it doesn't look like this quality here. Here's the one that I just made. Here's the quality of the one that was shown off on Reddit. This one, hard to tell that it's fake. This one, I mean just the colors of it look off. That skin's got like a little plasticky feel to it. This to me looks like an AI image. I don't know if I've just seen so many now that I'm better at spotting them, but there's definitely a quality difference between this and this. The one that I generated here came straight out of Glyph with no extra filters or anything special running from it. It's just the prompt into Glyph giving me this image. The ones in this example, on the other hand, they used what's called a Lora. Now a Lora is a low rank adapter. You can think of it as almost like a filter or plug in on top of the normal image generation. So Flux is the foundational IM model which generates the image. The Lora is like some extra sort of fine tuning information on top of that training. Here's how Perplexity explains A Lora Alora is used to train the model on specific concepts, styles or characters, allowing for targeted improvements in image quality, style specificity, or character consistency. Lora models are typically small in size, ranging from 2 to 500 megabytes, and can be easily integrated into existing models to enhance their performance. It allows them to customize their AI models to produce unique art styles or improved image quality without requiring extensive computational power or a complete retraining of the model. There's some examples they gave here. Style specialization training the model to generate an image in a specific style, such as anime or oil painting. Character specialization training the model to generate images of specific characters Such as Mario or spongebob, or quality improvements enhancing the overall quality of the generated image, such as improving the detail or texture. So somebody basically trained one of these Loras, which works in combination with Flux without needing to retrain Flux entirely. It can just add the additional information that's needed to get to the desired result that they're going for. So with this example here, they used a Lora from XL Lab, which apparently affects the skin, the hair and the wrinkles to make the images look more realistic. Same thing with this image here. It was using the same Lora to get this sort of extra realism out of the image. However, when I'm using Flux inside of Glyph, I can show you when I look at my actual Glyph workflow here, there is no special add ons here, there is no Loras happening. Even under advanced controls, we don't even have the option to add Loras in. And even if I click Add Block, there's no options to work with Loras within here as well. Something that I imagine Glyph will probably add in in the future if I had to guess. But right now we don't get that option. We get what comes straight out of the Flux foundational model without the benefit of using that extra Realism Laura that the people on Reddit were using. Now, one option to be able to use the Loras would be to use something like Comfy ui. You've probably seen a few of these Comfy UI workflows. They look like kind of spaghetti bowls with these lines going everywhere. They get complicated really, really quickly and are over the head of most people. I even struggle to wrap my head around them once they start to get more complex than, you know, three or four blocks. The other way to use this Laura, would be to use a site like File AI. Now, this is a service similar to Replicate or like what you'd get on Hugging Face spaces, where you can actually run AI models but you're using their cloud to run them. You're using like the fall AI cloud to run the inference to run the processing on these AI images. Now they have the standard Flux 1 Pro model here. So if you want to just use Flux Pro, you can use it. But we're going to run into those issues where if I want a super realistic image, it's not going to look as realistic as what we're seeing because it doesn't have that additional Lora information on it. However, somebody did add the Flux Realism Laura inside of File AI or File AI here. Now, one thing to note when you do first start using this site this fall AI, it's not free to run the inference and to use their cloud computers. It costs a few cents when you do it. So every time I run Flux over here on this website, it costs three and a half cents or for about $1 you can run it 29 times. Here's the thing though, when you first sign up, as of right now, as of the recording of this video, they actually give you $2 worth of credits so that you can get in here and play around with this. So if you do want to play with it yourself, you've got a couple bucks, you know, 60, 70ish generations before you need to start spending out of your own pocket. But once you're logged into this fil AI site, you can go to Fall AI slash models and see all the various models that are available to use here. And at the time of this recording, Flux one Dev and Flux Realism are both towards the top for you to test out and use. Right now you can also use the Flux 1 Pro, which I believe is slightly more expensive. Yeah, so it's $0.05 per generation typically. But let's go ahead and use the Flux Realism lora here. I've already got the prompt plugged in here that gets us a similar image to the dude with the ponytail. And under additional settings. One thing I did notice is that you can leave the number of inference steps at 28 here, but if you leave the guidance scale at the default3.5, it actually doesn't look that great. It starts to look shiny and plasticky and more unrealistic. I found the sweet spot to be about 2. I would get in here and play with it. If you don't get an image that looks like what you want, you can play with this CFG scale here. But two was the sweet spot for the realism for me. I go higher than that and it starts to look a little bit more fake. So let's go ahead and change that and click Run and we get images like this which are pretty dang realistic looking. The forehead is still a little bit shinier than I'd like, but it's pretty dang good. The next thing I wanted to do was I wanted to animate them like we saw in the other videos that were circulating all over X. So let's go ahead and download this image and we'll jump over to runwayml.com and I'll animate it with Gen 3. So let's go ahead and click started on Gen 3 Alpha. I can grab my image Here and drag it in the one that we just created. It wants me to crop it. So I'll go ahead and crop it like that, so his whole head is in the picture. And then I'm just going to grab the exact same prompt that I had originally pasted in here. It's a little bit too long. It goes past their 500 here. So I'm going to get rid of the last sentence. Let's go ahead and generate this, see what it gives us. And here was my first attempt with the video. Oh, the magic. The microphone did a little magic trick there, which just floats in midair. If that didn't just happen, it's actually pretty decent looking. I mean, it's a good video up right until that moment right there. I think it's a pretty solid video. And then after the floating microphone incident, it actually still looks pretty dang solid and realistic as well. I generated one more time because this first video that I made, I accidentally had it set as the last frame being this image. That's why the video sort of starts zoomed out and then moves all the way to finishing on the frame that we had here. That was a mistake. I meant to set it as the first frame. So I generated it again where this is now the first frame. And here's the version we get out of that. This one is a lot more realistic. The microphone is kind of a little too still for my taste. Right. He's moving a little too much in that microphone. It's just like solidly held there. No matter how much he moves his head or mouth around. Which is like, looks off to me, but it's pretty good. I mean, the fingers get a little wonky here. But this is really how they're making those videos that you're seeing all over X right now of ultra realistic but not real people videos. The other way to do it outside of using Runway would be to use Luma's Dream Machine. So I went ahead and plugged in my image into Luma's Dream Machine here. Give it the exact same prompt. And I mean, the results are not quite as good as what we got out of Runway. You can see the face sort of goes wonky towards the end. Like, what? Oh, my God, what just happened there? So not great. Runway, to me, did a lot better job. Unfortunately, the way AI is right now, I'm fairly certain that most of the videos that you're seeing on X. Where somebody took a ultra realistic AI generated human and made a video of that person speaking. It was probably a little Cherry picked, they probably had to do a few rerolls. Like if I generated two or three more times, I bet one of the videos that came out of it would be just totally perfect and hard to tell that it was AI generated. But I wanted to make a quick, fun breakdown and see if I can recreate some of what I saw. There are ways to get it way more dialed in, right? If you want to use one of those comfy UI workflows, you can get way more dialed in using something like that. But I wanted to try to figure out the sort of quickest path from A to Z here. The easiest way that I've found to do it so far is to go to this file.AI. use the Flux Realism Lora model. Make sure you have your guidance scale set to 2. But if you don't get what you're looking for, play with that one a little bit. Use that slider until you get something that looks the way you want it. Then you could take the image that it generates, pull it into Runway, and now you're getting that ultra realistic video. Mine don't look as good as the examples at the beginning, but again, I really, really think some of that stuff was cherry picked. Just some fun toys for us AI nerds to play with. Hopefully you learned about some new tools that maybe you hadn't heard of yet this fall AI site. Actually a new one to me. My buddy Angry Penguin who was on that Flux video with me last week is the one who pointed this site out to me. And now I've been playing with it and using this Flux Realism Lora to really dial in these realistic AI generated images that are coming out pretty dang solid. So anyway, just thought I'd shoot a fun video nerd out with you for a few minutes today and show you what I've been playing around with. Hope you enjoyed it. If you did like this video, subscribe to this channel. More videos like this will show up in your feed. And that's it. Really appreciate you. See you in the next one. Bye bye.