OpenAI Just Perfected AI Image Generation (See the Comparison!)

Whoa, whoa, whoa. So we just got a spontaneous OpenAI livestream and they unveiled their new image generation model. This is a capability that is present within ChatGPT on all tiers, including the free one today. And I think a lot of OpenAI releases are very niche and for a specific audience and not everybody is going to make use of them. This is not one of those. I think this is really one of those moments and one of these tools that, well, it's not just going to be accessible to a wide audience, but also useful to a wide audience. So that's why today, in this first look video, I'm going to show you various things. First of all, we're going to have a look at the generation and how it works, just to get you quickly into it. Then right away I'm going to follow up with some use cases. I mean, look at this. It can fine tune the model off of one image. I just gave it this and boom, turned me into a firefighter with a simple prompt. Then we're going to look at more examples at some of the special capabilities here because there is some like doing text like this. And then I have a bunch of benchmarking prompts that I already ran and we'll be comparing it to some of the top models. So here in this quick, concise video, I'll give you my first thoughts on how this performs, comparing to all the other models out there, like Imagen 3 and Flux and Midjourney, how to use this. That's where we'll start. And finally, I'll leave you with some thoughts on this model and on this release and how it slots in into the AI landscape. Now, let's begin with the first thing, the facts. I mean, this one is pretty straightforward. It's a new release inside of ChatGPT that rolled out immediately, so you can access it now. It's accessible to everybody on the plus Pro Team accounts and also on the free account. So if you just have an account with OpenAI, you can use this thing right now? Sure, there is restrictions and those always change over time. So I kind of shied away over time. I shy away from like mentioning those because they will update them. But you can just log into ChatGPT and use this. Now the way you use this is as simple as it could be. You simply say create an image or generate an image off and then you say whatever you want to do. I usually do a cat with a hat, as you might know. And then as you hit enter, this will start loading in. It is slower than the Previous tool that you might be used to, which was Dall E, but that was at this point, one of the worst models on the market. And this spoiler is probably the best one in the market. We'll talk about the nuances of that in a second here. Okay, There's a few things to talk about there, but as you can see, it's getting started. And within about, I don't know, 15 to 30 seconds, this will start rendering. It also renders from top to bottom, as you should see, any second here. Okay, so what makes this model special? Why should you care? Isn't there a thousand image generation models out there? Why is this one better? Why are you making a quick video about this? And why are people talking about this? And why is there going to be a thousand threads on Twitter tomorrow of 10 game changing ways to use the new OpenAI image generator. Oh, the new OpenAI image generator, you know, made all designers obsolete. Like you're going to have all of that. Oh, and as you can see, huh, it wasn't able to complete the request, so maybe it has a lot of traffic. Anyway, where was I? I was making fun of Twitter. Fred people. Fred boys. Right, okay, so let's get back to, let's get back to the content here. So what can it do? Well, it can generate images. And not just that, it can also edit images. So, so let me show you this quick example here that I did right away. Generate an image of a cat of a hat. It generated this beautiful Image in about 15 seconds. I mean, hyper realism on point. Okay, but there's many other models who do that. What else can you do? Well, one thing that you could do is you could selectively edit things. This is something they didn't talk about. So I could do this and say change the eyes to be red. Right? So you can do this, send that, it's going to update it. What else can you do? Well, you can edit the images. Okay, so this is one of the main points of this video. It's not just image generation, but it's also image editing. So this generated an image, but now I have an image and I want to keep working with that. So I can just follow up by saying change the hat to something a nobleman in the Middle Ages would wear. I'm playing Kingdom Come Deliverance right now for anybody who knows. So that's why this is inspired by that. Anyway, here you have a noble cat from the Middle Ages. And then I could go on, add a shirt with funny text referring to cats. So, you know, I'm using the power of the large language model. This is not like interacting with midjourney. This is not, hey, I want exactly this text or I want exactly this type of result. I'm using GPT4O's capability to come up with a funny text. And then this image generation model seamlessly integrates. Sorry, I can't. I have plans with my cat. Perfect. It's a great shirt. Color correction has been done and as you can see, it works. And this is like the big points they led with. But enthusiasts might beg to differ. But they're saying like this model can do flawless image generation. But we sort of had that before. No, and this is where my comparisons table comes in. Because, yeah, if you look at some of these other models like Imagen or Flux Pro or even Image Journey now can do short text. Well, long text, not so much for long text, you might want to go with ideogram. Sort of hit and miss, but it works. But you. But if you do really long text, most of these will break. Well, not OpenAI's. You can do things like this. You can do a ticket with multiple paragraphs of text immediately. And it gets it right every single time. Not most of the time. Not sort of. It just works as you can see here. So that's pretty amazing. Now this is just one of the capabilities. There's a few more. I'll quickly name those and then we'll move on to more example more examples. So another one is that it can work with multiple files. This is also something we haven't really seen out of any other model except maybe Google's competitor here. And just here I want to show you. I ran the same thing, right? Turn me into a firefighter with an image of me. This is what I got out of Google. Where's the ChatGPT one? This is what I got out of ChatGPT. It's not even close, right? It's not even. This is absolutely terrible. And this is absolutely incredible. Now you would get something like this out of flux if you trained it on a few of your images. But you wouldn't get something that this closely resembles from one image. Let me tell you that I've done many fine tunes. I have many fine tunes of myself. We use them in the thumbnails. As you might know. This just works. So I don't know wins out on that. Also, what it can do, it can do transparent backgrounds. So remove the background and turn it into a png. I'm going to be extra precise here. So it works here for sure. But I believe just by saying remove the background, it will actually cut it out and turn it into a PNG file, which you could then copy and paste anywhere else. And if you're not familiar, PNG files are image files like JPEGs with the background cut out. Okay, what else? Well, we have a few benchmarks here at the AI Advantage because as you might know, we do a monthly ranking. So if I just go into the public area of our community here, then this is accessible to everybody for free. You can just review this. We sort of do a ranking for image, video and LLM platforms once a month. This is the version for March and it doesn't include Imagen free yet because we include things that are like, publicly available. And for Google AI Studio, we just got it now before it was in, like, early release. Anyway, the point is this is a ranking of all the tools, okay? And Imagen would belong in the S tier. So now the tools really to beat here are midjourney, flux and imagen3, which is this, by the way, which also has image editing. Midjourney and Flux only do image generation. I mean, I suppose you have editing tools on top of the stuff, but you can't really do things like this. You can't upload an image and kind of work with it in multiple steps and expect Photoshop like behavior. That's what I mean by editing. Okay, so how do these models compare? Well, we have a set of test prompts and I ran all of these for this new model. So what I want to do now is look at these different test prompts. Here you have them. Logo design, portrait photography, cinematic still, aerial photography, book cover and comic book. It's a total of 1, 2, 3, 4, 5, 6 prompts. And I ran all of them through ChatGPT. That's these tabs up here. So let's. So let's have a look together and evaluate it. But before we do that, I just want to kind of look at the image of the fireman. This is still creating to give you a feeling for, you know, how long it takes. And then also this one is still creating. So I suppose now it will be a little slower because, you know, everybody's catching on, everybody's using it. I ran the test prompts right after the stream, but as you can see, it's slowly, slowly generating here. I'm sure this will work well before we get into the test prompts. And I forget, actually one more thing. It can do things like this. You can give it your brand guidelines, your colors, fonts, you want, and it will actually use exactly those colors. Now, this looks funky, but I gave it different shades of green and a purple that we use and a gray, and it used all of them. And also it used the prompt I told it to. So you can fine tune it on one image like this. And look at that. The firefighter is coming out. Great. You can give it one image and do images of yourself in all kinds of scenarios. You can cut out the background. You can use it for marketing purposes with things like specific colors and fonts. And now, to round out the video, let's look at how it compares to all the other tools, because that's what everybody's curious about. Well, I can already tell you it does super well, but the specifics matter. Okay. And the specifics are to be determined by this comparison. So I'm going to look at mainly these. Okay, maybe there's Ideogram two over here. There's a few more Canva Magic Studio and stuff. But really what I care about is the Midjourney column, the Flux Pro column, and then maybe Recraft, Imagen and Ideogram. But sort of these are the S Tier tools and also Imagen, these are the tools that I really want to kind of review this against. So we're going to start with this logo design prompt. Okay. Logo design. This is the way the others look. And what I'm going to do here is I'm just going to kind of pull this to the bottom so we have a nice view of both. Okay, There you go. So for Logo design, you can see this is what we got out of. This is what we got out of ChatGPT 4.0 image generation. That's what this thing is called. I would say it's very simplistic. You know, it probably is most similar to Imagen Free, what imagenfree did here, but even that has more shading. So I honestly like. On Logo design, I think Recraft is kind of the best here. And Ideogram is also very interesting. Like, there's a specific style over here. Can you. You can see it here. I'm gonna make sure that my image doesn't cover it up. But this is Ideogram. I think this is probably the best one. Or imagenfree does a really good job at a minimalistic look. But this is just simple. There's no shading. I would say it's slightly unclean in the lines. Like, there's an extra kind of. What is this? If I got this from a designer, I would be happy with this. Look at that. This is not clean. This is not clean. This is not clean. So I'm not a big fan of the logo generation here. Okay. Just straight up. It did the text perfectly though. And if you need a quick logo, there you go. But for that specific use case, I would probably prefer another model. Let's look at our next one though, which is portrait photography, a very important one. Okay, so portrait photography. Here we have the exact prompt ran through ChatGPT 4.0 and this is a fantastic looking image. You can tell already? I mean, I mean, you can really. How do I max this out? I think this is actually the best view you can. Really the same. The skin texture and the pores are beautiful. The suit looks hyper realistic, just looks perfect. The beard a bit too perfect actually. And if I compare it to all of these, well, hyperrealism sort of was solved with these models. Right. Flux Midjourney, both doing an excellent job. Midjourney always has this slightly more like artistic and artsy look. Like a creative director created the images. Like some, you know, very tasteful and avant cards creative director created them. That's, that's sort of the thing that you get with Midjourney. You get these like, tasteful, artistic images. Flux is just super good in Hyper Realism. So this one is really hard to judge. To be honest. I think Flux Ultra might be slightly worse, maybe, but not, not really. I don't know. Honestly, this one is super hard to judge. Recraft is definitely worse. Yes. And Imogen, well, it shows a bit of a different individual, but that's always going to change slightly depending on how often you prompt. But in terms of Hyper Realism, I would say, well, it's above Recraft and it's on par with Flux and Imagen. No real difference there. It doesn't matter. Well, I guess it matters because this is inside of ChatGPT and that's what most people use, so that's an upside. But other than that, in terms of quality, I think Logo was worse and I think image generation is on par. This, you know, common people who don't spend their, you know, where AI is not their hobby wouldn't be able to discern this from a normal image. Okay, so let's move on to the next one, which is a romantic cinematic still right here. Shout out theoretically media for originally showing this to Matthias in our community and Matthias using it all the time now. And now we're all obsessed with cinematic still prompts. Anyway, that's a good, good little prompting trick from, from fellow YouTuber Theoretical Media Point being, we ran this prompt through all the models and you can check out the differences here. I have one comment, and that is that I ran this prompt twice. I had to run it twice because in the first time it refused. Okay. So it didn't want to go along. And it told me, hey, I cannot do romantic scenes. So, you know, it's pretty. It's not super restricted from my experience so far, but there's definitely some restrictions. Also. This is just the first look, right? Cinematic still. How does it perform? Well, let's look at Midjourney. It does these extremely film like sequences. There's. There seems to be, you know, like the. The blacks are pretty gray. So like, there's a very distinctive look to it. It's sort of like it has like a filmish color grade. And I would even say maybe there's a Pro mist filter on it or like some. Some vintage lens. Like, this is very stylized. That's what I'm trying to say. This one is very similar to what we get here. And you know, honestly, how do you. How do you even, like, judge this? It's just different flavors of good at this point. Right? Like, honestly, if you look at these two, Flux 1.1 Pro and Flux 1.1 Ultra results, I mean, I guess these look more like Polaroids with a flash, but this looks like a movie scene. I mean, this one is obviously worse. So Recraft is worse at these again. But what about Imagen Free? Well, Imagen Free is excellent. Maybe the motion is a bit stronger here. This looks more like a poster, I would say, but I mean, just quality wise, they're all on par. They are all hyper realistic. There you go. I got it out. So maybe one last comparison here to Ideogram. Ideogram. Just looking at these images briefly, the quality of these is lower. But as you can see, Imagen Midjourney and Flux all sort of like on the same level. With these comparisons, it's really hard. Like you can make up your own mind. You can look at this video, you can pause it, you can zoom in, but they're very similar. So let's have a look at the next one, which is a aerial image of, well, Mother Nature. So let's have a look and. But I can already tell this will be very similar to the others. It's on par with Flux. It's on par with Midjourney. Midjourney again takes more of a cinematic approach. There's more clouds, there's. There's a stronger color grade. They really push the saturation the blues. So again, it's just more cinematic. And then recraft here just didn't perform well at all. And Imagen sort of gives something that looks like a. Like a photoshopped background, I would say. So in this case, I think Imagen is actually worse. This just doesn't look real. This looks artificial. Like, honestly, like this. Like what kind of. This looks like something like a kid draw and then they turned it into AI image. So I'm not a fan of that. An ideogram also doesn't work. But again, it's sort of on par with Flux here, right. In terms of realism. So fantastic job, though. Like that being said, it's on par with the best in the industry, right? In these things. But again, I have to point out, Flux doesn't allow you to give it one image and get a firefighter image like the one we looked at. And. And this level of quality, Flux doesn't allow you to cut things out into PNGs at this level of cutout. Flux doesn't allow you to do long text flawlessly. Right. Flux isn't integrated into GPT4O that you're probably using anyway. So those are big advantages. So that's. So, you know, the conclusion is probably, if you have the choice, like, just use this if the quality is on par. I just really wanted to look at the quality because that's important. Two more examples. So one of them is over here. A book cover. A book illustration with the text, the forest. Okay, let's have a look at all the examples we have here. And I hope. I hope this is large enough. I think it should be okay for comparison's sake. But there you go. We have the forest over here, and with the book covers over here, we have Midjourney looking really good. I think Midjourney does this really well. Flux, on the other hand, I don't know. This font is kind of corny. I don't know. I'm not a big fan. This is just too basic. This is just not good. Okay, but like, this poster is incredible right here. Also, this poster. Those look really sold it. Sold it very well. Just like Midjourney did. Just like this. Did it. But I would even say the midjourney one, sort of like the artistic flair again, it just has that on top of it. So if you need, you know, if you would want to hire a creative director for the image generation, or if you would want somebody who's very creatively involved and want something that stands out, I gotta say, I think midjourney still takes this home like this direct comparison, but it just doesn't offer some of the extra functions, you know, like the long text and PNGs and the things we talked about. Okay, what else? What else we got? Then we have the models that are really good with text recraft here. It goes with like a minimalist style in these cases. And this is more of a photorealistic one, honestly, just different flavors of the same thing. I don't think one of these is better or worse. And then lastly, what is this one? It's imagen free. Oh, yeah. This is what I'm interested in. It picked a very similar font to this. The image is essentially the same. Honestly, this is just a darker version of the image and thing. Like, they just look identical. So again, in all these cases, look at that. All of them did the text flawlessly because it's short text. Right. And then we have one more example, the comic book that I want to move on to, and that's this one. So let's have a look. Beautiful, Beautiful work from GPT4O right here. Look at this little comic strip. I'm sure it could add comic book text. I don't know. Super. You know, I'm recording a video. I'm not going to come up with what exactly to say here right now, but this is just beautiful work. And it creates the entire strip. Look at that. Like, that is some advanced stuff. It doesn't just create the image for you, but the entire comic book strip. Because, well, it should be a strip illustration. And in the other models, it disrespected it. Look at that. These are just images of the woman looking good. Looking good. Also the flux ones looking good. But none of them really feature the hero in a dramatic way like this. Maybe these ones do. What is this? Imagen free. Imagen free. These ones do. There's also text on them. Kind of makes me want to. Want to increase the size on this and really show you the comparison with imagen here. Let's do that. But yeah, let's have a look. This is the one on Imagen. Yeah, the text is all messed up. Let me tell you. Like, some words are correct. Like, even if you can't read this, I can read it on my screen here. Some of the words. Let me zoom in more. Some of the words are screwed up. Look at this. T Y I s your adnitz Anne. Hey. Brio. Yeah, it's just not coherent. Utah you thielo. It's just not correct. Okay. So other than that, I think this is really close and good. But on this 1 I think GPT4.0 clearly takes this home. Wouldn't you agree that this is just the best looking one out of the bunch? And when I say add comic book text, oh, it gives me ideas on what text to add. So here you can work with the LLM. So as you can see there's real advantages to this. On all the other things it's equally as good or maybe maybe with the logo slightly worse. Arguably maybe. But just the upside of having GPT4O at your disposal and you know it telling you exactly what text it came up with. The LLM, you don't have to round trip to ChatGPT to generate the text. You can just do it all in here and now I suppose it will do this like that is super strong. If you add on the capability to, you know, use that LLM like this and to use the selective editing tool like this to just change the eyes and maintain everything else right from the first image. To be fair, it did now change the background to the brand color I gave it. So you know, if you wanted to do this cleanly, you would just generate the image, use the tool and then regenerate. This already took in the context of my further conversation, but it did the eyes as you can see it. Well this is still working, but you know, to be fair it must be overloaded right now as it just launched a few hours ago. But this is working, it's going to work, right? You're going to have a PNG that you can download and again these are all the capabilities you will not be able to get generate the image, you will not be able to get with the other models because you have all of the tools combined, right? So really here it's not just about the image generation, it's also about the image editing and it's also about the fact that you have a built in LLM and I think if you combine all of those things and you give it away to everybody on the free tier, well then again a master strike by OpenAI and remove in kind of like out competing the competition here again. Now this makes me really curious if that's true across all use cases or if it's just on this one quick little example. And that's why we're already creating a video which will be comparing various use cases between Google's image in here in Google AI Studio that we pointed out on the channel about two weeks ago when it released the news you can use show and OpenAI's new image generator. Because this is really good at editing. But OpenAI's image generator is really just not just good at editing but incredible editing. It's kind of perfect at editing. It can do long text which no other model can do and all the other things I pointed out multiple times in this video. So there you go. That's your roundup and comparison versus all the other models. Let me know what you think. Are you going to be using this? Are you going to be trying this? It's really simple to use, right? Because it's just on a free account. You can try it out right away. I think this is sort of a. Well it's a. It's like a. It's like a big part of photoshop built into ChatGPT now that you can just activate within your conversations, can add multiple images, you can edit them together, you can alter things, you can cut things out, you can train a model on one image. Impressive. Impressive. OpenAI and good job. I kind of outplaying Google here again. So that's really everything I have for today. Subscribe if you want to see the comparison video versus Google's tool with a bunch of use cases for this and other than that, I'll see you on Friday in the news you can use show. That's all I got. Don't forget to like and subscribe. Bye bye.