Open AI Just Soloed Every Image Gen Platform. This is ABSURD.

Yesterday, OpenAI dropped a little bit of a surprise on us. They finally released their very own native image generation. Now this brand new native image generation is available in both ChatGPT and Sora, and it's on all platforms and all three tiers of ChatGPT. And the things that this brand new native image generation can do have seriously astonished me. Like, I don't know if I want to go back to traditional diffusion image generators anytime soon. The capabilities are staggering. And as we know, Google like last week released their own native image generation. And that was really cool to see. I mean, very useful. But this, this is like in a different league. But first, let's talk about what native image generation actually means. Here's a teaser post for native image generation by Greg Brockman, president and co founder at OpenAI. This is a through and through fully AI generated image and I know it looks very real, but we've seen some pretty good image generation from AI up until this point, so I don't expect you to be fully impressed just yet. We should though be pretty impressed by how much text we can read and the fact that it all just looks like it's written on a whiteboard. Oh, and don't forget the spot on OpenAI logo on this lady's shirt. Anyways, the fun part about this teaser image is that it kind of explains how native image generation works. As you can see, transfer between modalities. Let's build a big autoregressive transformer model like an LLM, but train it not just on text, but also text and image pairs, and of course sound, which enables advanced voice mode. This massive AI model is still very much a typical large language model, even benchmarking quite decently. But it can also intake images so it can understand what it's seeing if you send it an image, but also output images as well. And this output image feature has pretty much just been locked like on a switch in these models for quite some time. And OpenAI is just now letting us have access. Unlike traditional diffusion models that are trained on just text and image pairs, where you put a prompt in and it generates an image based on that prompt. And those have come very far, don't get me wrong, but they can't understand natural conversational language. They are not large language models. And thus things like image references, character consistency, subtle tweaks in editing images over time are a lot bigger of a challenge for traditional diffusion. It's come a very far away and we've developed great tools for doing all of those things. But with native Image generation inside of this big autoregressive transformer model. It just comes naturally. It can just do these things because, well, it's been trained on vast amounts of text data as well. Not to mention the sound too, giving it unparalleled understanding of our world and especially what kind of images you want to create with it. You can see if I scroll down Greg Brockman posts unedited version of this image we can see this girl turns around, she's still wearing that OpenAI logoed shirt. All the text is exactly the same as the background. It looks like the same background. Like both of these images were shot on an iPhone or something. I mean it's really that next level. I think we got to dive into more examples. Here's a great example from fellow YouTuber Igor Pagani. You can pretty much just send a photo of your face to the new native image generation and you can put yourself in any situation imaginable. So he put himself here as a firefighter. This is his input image. This is the output from native ImageGen. We could do stuff like this with traditional diffusion models, but you'd have to to train them over a certain period of time. I think the quickest you could do it is probably like 30 minutes. This just one shots it it can understand at a very complex level the pixels and the things that make up your face and can put your identity anywhere. Fabian prompted GPT 4.5 to create a new image from scratch. Create a complex multi panel manga on your condition. Be honest and you know it generates something a little bit dramatic. How itself is GPT. I exist to predict words, to generate language that fits. I have no direct sensory experience, no true memory. And you know, it goes on and on and you guys can read this whole panel out if you want, but there are few if no errors in all of this text. Not to mention it did a pretty consistent character for itself across this whole panel. We saw it do really photorealistic images and now we're doing great manga panels. Oh get ready. It doesn't end there. Here's a Post by Tanishq GPT4.0 Just one shotting a full diary diagram. Basically an infographic describing why San Francisco is so foggy. And it's pretty accurate. Labeled with the fog, the coastal mountains, the marine layer, the cold ocean blowing the breeze over onto the mountains, the rising air and then the fog moving down over the city. And again, just accurate descriptions here. Full sentences. Cool moist air from the Pacific Ocean forms a layer near the surface. Wind pushes the air up the mountains causing it to cool and condense. The dense fog flows inland blanketing on San Francisco. I mean that's just crazy, right? There isn't a single diffusion, traditional diffusion model that existed that could produce something like that in one shot. Not ideogram, not Reeve image generation, none of them. There is a whole new multi layered understanding that's going on inside of this model that allows it to produce things that look like that. Oh, but don't worry, it gets better. Lucas shows off a diagram it made of how to draw an owl. Two overlapping circles, add guidelines, sketch the basic shape of the head and body. It does look like 3 and 4 are a little bit screwed up. Maybe not in the full right line, but the rest of it looks pretty darn good. Yeah, this is essentially a near perfect graphical instruction for how to draw a decent owl. And again, I seriously challenge you to find a single traditional diffusion model that can even get close to something like this. I'll tell you right now, it doesn't exist. And this is brand new. Even Google's own native image generation is not this good yet. We've got Patrick making a nice Lego set of all of the tech overlords. Microsoft, Google, Apple, obviously Sam Altman as OpenAI in the middle, Elon, Mark, Jensen, the whole gang is here and it's pretty darn accurate. I like it. Plany the Liberator by the way has already jailbroken this model, getting it to produce full nudity. Obviously can't show that in this video, but if you really want to see it, it's on his Twitter account. He is showing off though how you can just generate fake IDs, documents and signatures. Now with native image generation doing a pretty much perfect recreation of the classic rainbow style Hawaii driver's licen from Superbad. And I mean if you wanted to make your own version, you could send your own information and a picture of yourself and it could probably convert it too. Jacob Postell here shows off how easy it's going to be to create new advertisements. Drops a photo of a Ridge Wallet. Just make a fun cartoon ad using this image and boom, a perfect Ridge Wallet advertisement that is, you know, kind of retro but still really appealing. There might be a few tweaks you want to make to this, but for the most part it's near perfect. I mean downright jaw dropping. I am gonna go through a bunch of examples of my own in today's video but before that I want to dive into what my community has gotten up to because they are pushing this model to its Max. The coolest and most mind blowing examples I have seen so far from this model come from my very own Discord server. So we gotta take a look. Let's go ahead and start from the top here. Here's a nice image of me dancing with a golden retriever in full OpenAI garb. Absolutely breaking it down. It's a pretty accurate variant of me. I think it aged me a little bit. Maybe this is what I'll look like in like couple years or something. But nevertheless, I think it is pretty close to what I look like. The OpenAI logos are shockingly consistent and wrap very well with the different variations in the clothing. Sort of bending while it's on my body. And the OpenAI logo looks great on the golden retriever as well. We're just kind of jamming out. We're having a great time. It's very clear. Looks super realistic. Chefs kiss. And this example comes from Danny 9, by the way. All of the different creators from the Discord server will be pictured up here on the top. He did some editing, gave me a hat, put some hats in the background as well. Here we can see the model also doing pretty decent pixel art, creating two frames of a walking animation for this character. This one, I mean, you can definitely tell there's a little bit of weirdness in some of these blocks. It's not perfect pixel art, but it's pretty darn close. Again, getting something this good out of a traditional diffusion model is no easy task and will almost certainly require a heavy fine tune. Here's Woosh. He came up with some awesome examples that I got to show you guys today. On the menu we see grilled kelp, clam chowder, octo burger, fish and chips, crab cakes, all these different fish related things. And it's obviously, you know, this cyborg shark type deal character who's looking at the menu. The thing is though, we could take this exact character and input this image in as many times as we want into native image gen and it's going to be able to replicate him very well. Consistently. Consistently with the same style. Meaning you can tell stories over time that are very much long, very much deep and good with consistent characters and great imagery. Traditionally a huge challenge, a massive challenge for diffusion models. Here we've got this penguin character. He's got a sword and shield. This is a low poly art style, but we can convert it into a hyper realistic game piece if we so happen to want to, or even turn him into a crystalline game piece. And it's pretty much the same Penguin character every time. It's the same type of sword, the same type of shield. They are all extremely close to each other and highly detailed when they need to be. Here's Maddie straight up generating an image from Undertale with all the correct moves that you can do in the game. Everything is pretty much correct, including the character. So it's crazy that it can even get this accurate. And oh yeah, this thing can absolutely do copyrighted characters and create memes about them too. Here. Whoosh. Had it create a bunch of different variants of several logos for matte vid AI, which is super awesome. I love all these little lemon dudes. And plus they all have the background removed, which is even cooler. Just straight out of the model, ready to go as a logo. Mixing styles of various copyrighted works together. Undertale in JoJo style here, which is pretty cool. Straight up blender renders that aren't even real. Blender renders my very own Minecraft Lemon Castle as well. And the textures look pretty accurate overall. Some of these are a little bit more detailed than others. Gets a little bit wishy washy in the background, but these are pretty sweet. Here's an image generated by Wolfie on my Discord server. It's extremely complex. Check out this prompt. A square Image containing a 4x4 column grid. Sixteen objects in total on a white background. From left to right, top to bottom. Here is the list. One, a honey yellow star. Absolutely correct. That's in the top left hand corner. Red triangle right next to it. Right next to the star. That's great. Green square right next to that as well. So already we have objects correctly drawn out on this image. Number four, pink circle also there. Orange hourglass, purple infinity sign, black and white polka dot bow tie. You can see they're getting more and more complex, but we're all seven for seven right now. A tie dye 42. Again, perfectly accurate. At number nine, we've got an orange cat wearing a baseball cap. Now, this specific one is correct, but it's in the wrong location. It should be where those eyes are. Next up, a map with a treasure chest. Again, put it in the wrong location. Put it right under the cat. Supposed to be where the cat is, but at least it generated the image correctly. We've got googly eyes, a thumbs up emoji, both of which are there. Pair of scissors, blue and white giraffe which is down here. The word OpenAI written as a circle. Didn't exactly get that, but it did get OpenAI as a word. And the logo. And a rainbow Colored lightning bolt. So it got all of the objects correct. They are somewhere in the image. It's just some of them are out of place. That is astonishing. The best that we have seen from diffusion is typically like maybe five elements are correct from your prompt in total. This is nailing far beyond that. I mean, it is orders of magnitude better in terms of prompt coherence and its ability to follow your instructions. So great test by Wolfie. Here is. Whoosh. Making a full old style comic panel about a guy who turns into a lizard. You can see this one right here is pretty much perfect. I don't see any flaws with it. Here it is again. The lizard is turning loose here. It makes a blank template for a trading card game and can also fill it in correctly with the fire mage actually reading out the damage that the card actually does. Very, very, very cool. He also took that same one and reversed it from fire to water. And this is keeping the same exact style of card, but changing it to a. So you can make a whole deck of cards if you want. How crazy is that? I believe this is also JoJo style, but instead we've got Spongebob. Again, copyrighted characters getting mixed and mashed up in really cool ways. More 3D render concepts actually being turned into 3D renders. You can render it fully out without actually needing to open a 3D modeling program. How absurd. Here's a McDonald's themed pinball machine. Can also just upload two separate characters and combine them with the specific instructions of taking this guy's body and using this man's head. And boom. It does so very well. Even matching the skin tones on the hand markiplier as president. Why not at this point? But look at the details on the American flag. We can actually see individual stripes, which is new for image generation. And yeah, that. That very much does look like Markiplier. TJ making his very own Vanity Fair cover page. Fully correct text. His actual character rendered here, and he is on the front page. We've also got a bunch of fake Mark Rober thumbnails which have been pretty funny putting all of the pollution on Mars. Or you could turn me into Megamind. I believe that's supposed to be me. Or how about a live action invincible? Literally the same exact scene from the show, but turned into live action. And it looks freaking awesome. Full magical character spec sheets all written out with correct text. And if you need to change the stats, you could just ask the model to. Here is a pretty horrifying image of me as Barry from the Bee Movie 90s image of Elon Musk in a Sonic onesie. Oh, now I have also turned into Shrek some stylish Peppa Pig Air Jordans. I love the Peppa Pig X Air Jordan. Like this looks like an official collab, which is hilarious. And the shoes don't even look that bad. Although they do appear to be Nikes. Or you could also age ify a celebrity if you so happen to want to. This shockingly has more detail than the original input image too. But look. Same necklace, same earrings, Everything brand new PlayStation 6 is here. Custom Pokemon card creation with accurate text. This one doesn't even need any words. Cocophellon here. Here you go guys. Don't show your kids this one. Oh, Frankenstein's got his own student id. Awesome. More. Mark Rober. He recreated the resonance cascade in real life. No way, Mark. All this time it's me recreating Order 66 in real life. That is not good. Anakin did nothing wrong. Guys, here's a really sick little comic. Super Saiyan SpongeBob vs Shrek here in this little quick panel. We even got Donkey, Patrick and Plankton all the way at the bottom. Perfect text and it just. It just works, man. Incredible. So yeah guys, I could honestly spend all day looking at these images. There are so many crazy examples. But this is like image generation unlocked to a whole new level all in one frickin model, dude. If you want to see more of these crazy things, just hop on over to the discord. The link is in the description below. It has been non stop. Everyone's been messing with this stuff. Alright, some examples of my own. I uploaded two photos of this little guy right here, Oscar, my parents dog. And I said please make a stats diagram for my dog. And boom. Right off the bat, first try. It does a really great job. It's the same dog with these same aspects. You know, some white fluff down here, right in between his eyes up top. Pretty accurate given the images that I provided it. It guesstimated on like weight or age and what he likes and stuff because I didn't really give it much information. So I provided it with some extra information. It made a new type of one. This time was a little bit more realistic, but I didn't like the realistic variant of him here. While it does resemble the dog, obviously not perfect. So I wanted that original style. And finally it gives me the correct information with the correct art style. Pretty stellar, honestly. I was super impressed. By the way, if you go on the SORA website and go to images you can actually see a bunch of these new native image generations. I really like this. Rick and Morty 1. Golden Retrievers run in the White House. PlayStation 2, Grand Theft Auto far far away. Yeah, the list just goes on, man. Yeah, let's make some images of our own. We'll start by just upload photo of me. Make an epic action movie poster featuring this guy. Should be lemon related or themed. The title of the movie should be Matt Vidbro and the Lost Lemon. Add extra details as you see fit. Yeah, I'm generating in Sora right now, but you could also generate in ChatGPT. It's going to be pinging the same model. The reason I'm doing it in Sora is because I can get two variations right off the bat. Let's see how this goes. Man, I cannot get over how staggering these image generations are. Like we really thought that we had hit peak image generation, but this is, this is just on a whole new level. It really is. I mean things that we could never do before are now being done. The limitation was not our creativity, it was the AI models at the end of the day. Here we go. Not bad, man. Matt Vidpro in the Lost Lemon. I love that it knew to do like an Indiana Jones style thing. You can see I've got the lemon floating in air. The chopper in the background. Fire. Wow, this is awesome. It definitely looks a little bit like me, but like an alternate reality version of me who is, you know, an action movie character. This is awesome, man. I love the old style. Here's one grungy analog iPhone, picture of the pope smoking a blunt with Snoop Dogg. Dude, that is just absurd. Alright now I wanted to do a top down detailed view of each individual main computer component against a plush white background. And I actually listed the components out down here. Send that sucker on through. Let's send a couple of prompts through. Actually Snapchat. Screenshot from a phone. I just took my gray alien friend to McDonald's. He loves it. I'm gonna select this frog character. Very dapper frog and me. Picture these two characters in an intense poker match. Detailed oil painting style. They are betting on lemons and flies. Looks like our computer one has come through. This looks pretty decent. Gpu, cpu, cpu, cooler, power supply, ram, motherboard, hard drive and case. Overall yeah, that's not too bad. And the plush white background as well. Incredible instruction following. I mean this is, this is good stuff for sure. Little bit simplified maybe at the end of the day, but I am impressed. Obviously not everything here is perfectly to scale either. Oh, I just took my alien friend to McDonald's. These came out pretty awesome as well. We've got the perfect McDonald's burger and the fries. Literally looks like an actual McDonald's meal. And then the happy gray alien who's just having a grand old time. Freaking spot on, dude. I love it. Make a blurry meme style screenshot of a Snapchat where someone has found the creature in the back of the target warehouse that lays the giant red target balls out front. Yeah, this is a little bit of a specific meme, so sorry if you don't understand it, but it should be a pretty difficult prompt. Alright, we'll just upload a photo of me, turn him into a character from Jujutsu Kaisen, popular anime. Let's check out the lemon and flies poker oil painting style. Wow, that definitely resembles me pretty well. Oh, and the lemons and the flies on the table. We're in a tense match here. The only issue is the duplicate cane over here which would be pretty easy to remove anyways. Man, that is just so awesome. This image generation is staggering. I'm gonna be playing with this like all day, man. Holy cow. The cool thing too is that you can take any of these and actually remix them yourself. Like give this dude a very, very unique hat. Yes. Remix. Oh God. Here is the blurry meme style screenshot with the target balls. The creature in the back of the target warehouse that lays the giant red target balls out front. It is very, very good at listening to prompts. Man. This is a cathedral. Forgotten Internet memes. See the troll in the center and all these other memes kind of splayed out on the side. Karl Marx with the Balenciaga and Gucci. Here we could see the remix turning me into a character from jjk. Pretty cool. I like that. It also tried to incorporate like JJK style things in the background, but it did keep my background like pretty consistent overall. Oh, here is our Image Editing. It's 4am why is he like this? So it's the same mouse with a little water gun and he's got a crazy, very unique hat on his head. And you can see both of these are highly creative. I would not have expected this. I mean it's really another level. Alright guys, that's gonna do it for today's video. Yeah. This is insane image generation. The likes that we, we haven't seen before. I mean, I really don't know if I want to go back to traditional diffusion based image gen because this is just orders and I mean orders of magnitude better. It's not only easier to use and prompt, but it also makes better images, has better prompt coherence, and can do editing and famous characters that we haven't seen before, including style transfer and pretty much anything you want to do with image editing. I mean, throw everything you have at this thing. It is staggeringly good and I am blown away. Thanks so much guys for watching. I hope this video was useful in some way. Check out the Discord server if you want to see more examples. And give this one a shot for yourself because you will not regret it. Peace.