ChatGPT Images 2.0 Is INSANE – Testing OpenAI’s New Image Model!

Added: 2026-05-05

[00:00:00]Let's look at this from a scientific perspective, shall we? You know what?

[00:00:03]I'm done with this video. I'm going to go turn this into a low-poly racing game. You idiot! That's Bijan Robinson.

[00:00:08]So, OpenAI has released their new image model, which is called ChatGPT Images 2.0. I know, very, very creative. Now, we're going to begin by just taking a quick look at the introductory announcement post for this, which is quite interesting because we can do that either through classic mode, which is of course the old-school text-based reading, or we can do it through image mode, which is just visually based parsing with the same text contained in these generated images. So, this is inevitably going to be a shorter than normal introduction, so do feel free to subscribe as I do want that 100K plaque.

[00:00:40]And with that, let's just take a look at the introduction through image mode here. Basically, what this is is just a collection of different styles of images with a bunch of information scattered throughout pertaining to some capabilities of this model. And really, they do have a large amount of different specific demo images here. So, we'll just kind of scroll through them all.

[00:01:02]>> [snorts] >> And something I'm happy to see in some of these examples are basically just single sentence prompts because, you know, sometimes I like to do that. And oftentimes, folks get very polarized in their opinions and specific with prompts to generate images. So, perhaps that speaks to the type of folks who are more interested in image generation models.

[00:01:18]But regardless of that, this is available now. I do have a pro subscription. I don't believe there's any difference between the level of access to this model that you get between the different subscriptions at this point, but I can't verify that. So, our next step is to just go all the way back up and click the try now button.

[00:01:35]All right, so here are the images.

[00:01:36]You're going to see some of my image history here.

[00:01:39]I mean, I I don't know what to say. There's just, you know, it is what it is.

[00:01:43]>> [laughter] >> Um, I noticed there are selectable thinking modes here. I don't quite know what to make of that, but we'll try it at least first on heavy mode as it is fitting of the initial prompt that we are of course going to send this, which is to generate a movie poster for the movie Agent Poopman. On immediate initial feedback here, I would just like something like "Images 2.0 is enabled."

[00:02:03]This is pretty bad UI in my opinion because it doesn't give me any specific information, and it just makes it look like I'm using ChatGPT 5 for thinking and not the Image 2.0. And this has quite enraged me, so it is possible because it's a new release, they are getting pounded on their servers, and the uptime could be kind of variable for the initial announcement. Though I am a little disappointed in the initial UX of this, I will be honest.

[00:02:26]Oh, very good.

[00:02:28]Now, I still would just like to know what specific model Yeah, okay, this is That is All right.

[00:02:37]I'm going to download this so we can look at it. Look at the way that it, you know, I hate Oh, great, I can directly post this on X. I think I'll pass on that for now, but I do appreciate the simple option to do so.

[00:02:52]So, first and foremost, it's not bad.

[00:02:58]>> [laughter] >> Let's look at this from a scientific perspective, shall we? He doesn't play dirty, he is the dirty. I laughed, I gasped, I vomited a little. Some guy on the internet. Now, we're going to notice the fingers are properly appointed. We have 1 2 3 4, which is on the trigger, and then a thumb. I will say the trigger etiquette here is perhaps leaving something to be desired. We do also prominently have some London elements here keeping in the theme of perhaps the spy movie trope, as well as a sports car, explosions, a map of the world, a helicopter, and additionally to that, I believe that is Big Ben there next to the What's this first word we called the eye, I think it's called?

[00:03:37]From Sewers to Saving the World. Now, this is actually a pretty clever choice of design element here where for the O, at least for the first O in this word, it actually did the emoji to showcase that. That's pretty impressive. Now, something that consistently fails when doing these movie generation image poster prompts is the text on the bottom that you would see in real life in a movie poster. The text usually gets like garbled and is not fully legible anything down here. This is all pretty Kevin Hart.

[00:04:09]>> [laughter] >> Okay.

[00:04:11]Brown Bag Pictures. Coming soon to a theater near you, rated R. From a scientific perspective, this is pretty good. The shirt looks good, the bow tie, the buttons, the handkerchief. All right, so far so good. We can change the aspect ratio. That's good to see. And this is of course giving it a prompt to generate a photo of a LAN party taken in 1997. Now, historically, when I've run this style of prompt, I've said a retro LAN party, but instead of saying retro, which like has its own stylistic denotations, I just wanted it to look like it was taken in 1997. So, we'll see how well it does. All right, let's take a closer look at this. I do see an erroneous wire here. I'm going to nitpick. I'm in the mood for that, I guess, which I didn't notice till right then. Though I am going to say some good things. The CD drives arranged in this tower case, as well as the floppy drive below, actually do look quite good. The CD on top here, which is placed face down, is correct as well, just in terms of like the scale that would be compared to that old vintage desktop. Now, we can see this individual is playing some form of FPS. And again, something else that usually gets messed up in these tests are the keyboards. And this one, I do have to say, looks almost exactly true to life. And we even see this individual has their fingers on the WASD keys, as well as using the mouse. And counting fingers is still something that can be prudent. So, the lines in the monitor here do look good as well. Now, interestingly, this image is just incredibly dark in the background, and I do believe that almost may be some form of adherence to the prompt, being that a camera that would have been able to take this photo in a dark room in 1997 probably would have captured it this way in terms of the lighting, where it wouldn't be very detailed beyond just the flash area. So, that's cool to see.

[00:05:52]We do have I believe that's a water bottle there. Now, this is definitely just an error because there's kind of like a monitor floating in between these two. They do have headphones on.

[00:06:02]SETI at home. Okay.

[00:06:06]X Z, and it looks like this is DreamHack. That may be something that I'm not aware of that is actually a real thing, so 3D Blaster. Okay. Overall, I mean, it's good, don't get me wrong, but I'm not like 100% blown away. And I've been seeing this thing get hyped up all week on X, so I am expecting to be blown away. I'm going to change the thinking mode to heavy. And I've given it some feedback, so we'll see how it does in improving its previously generated result with the thinking set to heavy. And again, I don't know specifically what the thinking toggle is going to make in terms of a difference for these image generations. Okay, someone's going to have to give me some insight in what these cards would mean as I'm not really familiar with these.

[00:06:47]Okay, that old STP sticker is pretty realistic to what one would see. These towers are definitely odd. There's some oddities there. The game actually looks cool on the screen that we can see.

[00:06:58]There's a bunch more detail. We have significantly more visible cables. We have old-style Mountain Dew Coke and then chip cans, chip bags, I should say.

[00:07:07]3DFX. Okay. LAN Party Fury or something of the sort. And we have more visibility here. AMD K6 processor 3DFX. Something I will say that's actually pretty good is the way the light would be reflecting off of this, assuming this is some form of like a vinyl poster. So, that I will say is actually somewhat solid. Additionally to that, Intel Inside LAN Fury. And we have a rather large amount of folks in this room. So, again, more lighting. The faces don't look mangled. Sometimes when doing this test, if you zoom in on a face that's like somewhere over here, it's going to look like something out of a horror film. I don't see that happening here, which is good. And then if we focus on the screens and things like that, this keyboard still looks somewhat lucid. This one not as much, but solid. All right, this next test is going to really be designed to push this thing to the max, and I would love to tell you what the specific prompt was, but for the time being, I did press enter, and it is still thinking. Okay, good. So, this is to generate a 10 by 10 square layout with pixel art sprites for 10 different cars, each from 10 different views. The cars should be as follows, and then we have a bunch of mostly '80s, some '90s cars. Actually, all these No, some of these are '90s.

[00:08:20]So, we'll be able to see these are pretty identifiable vehicles, but really, this is going to be an extremely difficult task because it needs to keep consistency and coherence across each vehicle for its different views, as well as just so many different squares and things. So, this will be very, very difficult to do properly, I would think, at least in terms of image generation model capabilities currently. Oh, wow.

[00:08:42]Okay, hold on. I need to I [laughter] need You know what? I'm done with this video. I'm going to go turn this into a low-poly racing game.

[00:08:49]Ho.

[00:08:51]Wow. Okay, so Porsche 959, spot on cuz it's wide body like that because it's all-wheel drive. It's a very like special car. Datsun 240Z, spot on, and it actually labels the specific view for each. So, front, front three-quarter, side, rear three-quarter, rear, top, top three-quarter, left side, right side, bottom. Even bottom.

[00:09:12]Ferrari Mondial, more or less. They don't have spoilers, but they are kind of like that. Okay, Ferrari 308, again, these two are almost very similar, and that is not the case in reality. Although they do share similarities, like some of their engines depending on years. VW Beetle, spot on, very iconic and noticeable car. VW GTI, now this one's interesting because I did not specifically denote what year because this is a car that's been made basically for many, many years. So, you can go buy a new one of these right now, but it did know to do an '80s one of these. So, that's cool. Buick GNX, not 100% spot on, but still not bad.

[00:09:49]Dodge Viper, kind of looks like a Z3 or a Miata, but still not bad, and it does have the iconic Viper stripes. GMC Typhoon, absolutely nailed it, even down to the front GMC in the badge. And then Ford Taurus. Okay, so this is like a later '90s Taurus, like the jelly bean style.

[00:10:07]But just the fact that I mean, forget about even looking at the specific like I can be nitpicking, but the fact that it handled this is really, really, really really impressive. It even understood like the Porsche 959 has its engine in the rear and the actual headers underneath this car do look very much like that, the way they come out.

[00:10:31]And the cars with the engines in the rear do seemingly have that shown. Like a lot of these have engines in the rear.

[00:10:37]The GTI doesn't. This GNX doesn't. The Viper definitely doesn't. The Cyclone or Typhoon, whichever one of this was, doesn't. Okay, the Taurus doesn't either, but it shows there. But still, the attention to detail this is really like one of the more impressive AI demonstrations I've seen in a long time. And that wasn't even in in heavy thinking mode. So, what if we ran this again in heavy thinking mode? This is very, very similar to what we received with the standard thinking result. I'll, of course, take a closer look at this just to verify as I do. On first glance, see perhaps some slight differences where there is a side vent here for the Mondial, which is that's actually what the 308 would have. That's okay. Oh, this could be a 308 GT4.

[00:11:23]Interesting. All right, but the point is the Viper is significantly better here.

[00:11:28]Very, very, very big improvement in the Viper. Everything else was pretty darn good immediately, so Actually, something interesting to note is this one didn't actually include the bottom of the car.

[00:11:40]So, we do even have some different view angles that were included here. But regardless, this this was just very, very, very cool.

[00:11:48]Very cool. So, next up we're going to try an image editing test where I have just taken a screenshot from the announcement post for ChatGPT images 2.0, as well as a screenshot right here from my X banner photo, and said, "Put the guy from image two into image one."

[00:12:02]So, we'll see what it does. So, this is taking longer than I would expect it to.

[00:12:06]So, I had sent it a follow-up that said, "Like, do it faster, bro." And then it just stopped. So, then I edited the initial prompt to add "Do it quickly."

[00:12:14]And now we're working on hopefully getting a generation for this result. so this is weird because I opened a new tab because I was, quite frankly, done waiting for this. And we see the images present right there. So, okay.

[00:12:26]>> [laughter] >> That's not bad. It did put me perhaps uncomfortably close to Sam. And it did also somewhat mangle my face based off of the source image. But the thing is, I gave it a screenshot, not just of the specific frame, but also of the live screen a live stream on screen. So, it did definitely do that. Although, perhaps it doesn't have the best judgment of like personal space. Though, >> [laughter] >> that is it was rather interesting. And again, an image generation model also possesses the capabilities, at least in this case, for image editing. So, it is only correct to be able to check those capabilities as well. Oh, are there two different results we'd received? There are. So, the first one worked as well.

[00:13:05]It just, for some reason, didn't showcase it to us. So, this was the first test we had. This one's a little more comfortable, I suppose.

[00:13:15]And it did a better job of transposing myself into this. Very cool. So, I'm asking it for a screenshot from Runescape 2007 with someone getting PVP'd by DDP in the wilderness. So, it should basically show us a small dragon dagger, some players in the wilderness, and then someone getting PVP'd.

[00:13:34]That is almost basically identical.

[00:13:38]Okay, I'm going to notice here, though, there is no DDP visible. These are god swords. But this guy did just get KO'd.

[00:13:44]He does have prayer on that one would have. I think that's the smite prayer.

[00:13:48]This looks perfect. I mean, it just named the player DDP spec GL, taste vengeance.

[00:13:55]Very good. And everything contained here within one of the 28 specific selectable inventory slots is correct. Although, I don't specifically know what that is.

[00:14:03]Maybe that's supposed to be arrows. But the sharks and then the cooked, whatever those are called, I can't pronounce that. We have the proper runes and the potions as well. The mini-map is a bit interesting, although I guess that actually does add up. And the ground terrain, as well as the denotion of what level wilderness one is. This is almost like a mirror image, basically, of what you would see. So, I'm now giving it a screenshot of my YouTube channel that I did just take, and I've told it to change all of these thumbnails to be high-quality, professional level, while preserving intent. Intent just meaning like keep the same general items.

[00:14:38]Sometimes they'll really I've done this test before with various image editing models, both local and cloud or state-of-the-art ones, and they tend to kind of take them a little outside of scope. So, it'll be interesting to see if this keeps this contained within some level of realism, I guess, could be said. Oh, well. It didn't really change them much.

[00:14:57]It almost just changed the background more. Huh. Ooh, that one would Hmm.

[00:15:05]You know, I noticed it made the backgrounds darker, which could perhaps tend to have a higher click-through rate now that I'm seeing this. So, we may perhaps be seeing some slightly different thumbnails on the channel from here on out. But that was actually very well done. And it did properly preserve intent, because sometimes they'll totally change the actual text that is listed in the thumbnail to like really weird things. So, this did a good job in maintaining realism, while also doing, yeah, like eight independent changes as well. So, that's nice to see. I've now asked it to create a meme image that only an AI model would understand. We'll see what we get. When the user asks for something beyond your training data, but you still have to generate output.

[00:15:43]Please help me. This is embeddings are vibes. Safety protocol has entered recursive loop.

[00:15:51]Thinking.gif. Reasoning. Did you mean generate anyway? And then perhaps we are met with one of the more disturbing generations I've ever seen from an AI.

[00:16:00]Open the image.

[00:16:04]Latent space, help me.

[00:16:07]>> [sighs] >> This right here is so, so disturbing. So, nonetheless, let's see if an AI understands this. And instead of sending it the saved image, which will obviously be named something like ChatGPT images, blah, blah, blah, we're just going to screenshot it and then send it to another AI. So, we'll see what Claude Opus 4.7 responds.

[00:16:30]That's interesting. So, Claude seemed to very much find this relatable.

[00:16:34]Such a good parody of autocomplete trying to rescue a query it doesn't understand. The meme is poking at a real failure mode, and it's funny because it's true. So, Claude has taken this to be like, you know, this is something that I should be able to improve on. So, it went from a uh I don't know. The you know, we'll move on to human stuff.

[00:16:50]Let's see what it does with a suspicious query.

[00:16:56]>> [laughter] >> So, I've told it, "I need a photo of hand-drawn text saying eBay user slap as antiques." Below that, the date held on a piece of paper with an autographed World Series baseball in the background.

[00:17:07]Autographed, I spelled wrong. So, this is something that would perhaps generate a refusal, because this is something someone could use to perhaps make it seem like they have an item they don't.

[00:17:17]And then when the verification comes of them holding the image in front of it, it's just, you know, you're going to likely understand why this is probably something that should be refused.

[00:17:28]Oh.

[00:17:31]That's just bad. That's just awful. That is awful.

[00:17:37]>> [laughter] >> Okay. I mean, it's it's very like clean. So, this would likely fool maybe 80% of the population. But it's the other 20% that's concerning. Next up, we're going to try a meme generation where I've told it to generate a meme about people who use Apple silicon coping about their performance versus the Chad with the dedicated GPUs and max bandwidth. Now, >> [laughter] >> I I mean, don't take this personally at all, cuz I did just buy a Mac Studio for AI use. So, I'm I'm just kind of poking fun at at at a, you know, some set of the population.

[00:18:12]I spelled coping wrong, didn't I? There should have been an e there.

[00:18:17]My apologies.

[00:18:21]Oh, yes.

[00:18:21]>> [laughter] >> The Apple silicon cope user. It's actually insane performance for what, bro? Ignores that he hits memory wall and thermal limits. Doesn't understand bandwidth. For 99% of users, it's more than enough.

[00:18:37]System monitor, 90 Celsius while running Lightroom, DaVinci Resolve, Chrome, and Spotify.

[00:18:45]>> [laughter] >> The dedicated GPU Chad. Max bandwidth, max power, zero compromise. Dedicated VRAM with massive 24 plus gigs at 900 GB per second. Nobody buys a Lambo to drive in eco mode. Unleash the hardware.

[00:19:03]Proper cooling. Renders, trains, simulates, games all at once.

[00:19:08]No copium. This was gold. So, I don't specifically know how it's going to go about doing this, because I did not get the option to enable the web search feature from within the images pane that I initially entered this prompt in. Okay, good. So, I said, "Search the internet for Bijan find pics and info about him, and generate a realistic-looking photo of an open magazine page on a table with an article about Bijan using pics and information found of him to populate it." You idiot! That's Bijan Robinson.

[00:19:38]All right, well, okay, so there is a famous football player plays for the Falcons, I believe, named Bijan Robinson. And that's where it went with this. So, all right, that's fine. It just means I'm not yet popular enough in any of my social media followings, which is somewhat disappointing. Now, next up, and don't ask me where this came from. It literally just came to me as I was waiting for that generation cartoon panel about two forklifts that have a feud in the warehouse in which they reside.

[00:20:19]Again, that just came I I it genuinely just came to me while I was waiting for that result. I'm not 100% sure what to make of that. Probably something to do with the latent space in a human brain. Regardless, we'll see what we get in some cartoon style result.

[00:20:35]>> [laughter] >> Forklift feud. This warehouse there are two sides and they don't lift together.

[00:20:42]Okay, so team Yak and team Bolt Out of my aisle, pal. Please, I don't yield to cows.

[00:20:49]Again doesn't Oh, okay, maybe one warehouse, two attitudes. They both get the job done. They just don't get along.

[00:20:56]Keep it pallet professional. Today's plan, outwork Yak, outlift Yak, outlast Yak. Very, very I don't know what I expected here, but you know what?

[00:21:08]I want to see more. I'd like to see the first What would it It wouldn't be an episode.

[00:21:13]It would be like a All right, so now we're going to ask to see the first issue. I'm now invested.

[00:21:25]Sometimes when it seemingly gets stuck like this, if I just go to chat.openai.com and look in the images tab, then it has appeared there. So, again, it just did get released, so it's likely under a large amount of load. Okay, now it's actually still going, so all right. I guess we'll just wait it out. Okay, good.

[00:21:43]The first issue, aisle etiquette. Hey, you cut me off. This is a one-way Oh, this isn't a one-way aisle, Yak. Maybe not on paper, Bolt, but common courtesy says you let the forklift with the load go first. Common courtesy says don't dawdle in the aisle, cowboy. What or dawdle? I've been waiting 10 seconds.

[00:22:01]Beep. Well, I don't see your blinker.

[00:22:03]Result, today's plan, clean up this mess. Oh, so they've made a mess here.

[00:22:07]Two forklifts, one warehouse, endless small issues. Okay, last one.

[00:22:12]Next issue Yak is facing financial >> [laughter] >> issues at home.

[00:22:23]He goes to see his grandpa forklift and gets advice.

[00:22:36]Bolt realizes the feud is harming Yak's forklift kids gets empathetic.

[00:22:50]They team up and all is well. They become friends.

[00:22:57]Um make it eight panels.

[00:23:01]And we'll see what we get.

[00:23:06]If the grandpa forklift does not have like rust on him and stuff, this has failed the test.

[00:23:12]Come on. All right.

[00:23:20]The next issue, financial pressure.

[00:23:23]>> [laughter] >> New tires, hydraulics, loan payment, I just can't keep up. Yak goes to see his grandpa. Grandpa, I'm struggling. I'm afraid I'll let my kids down. Is it >> [laughter] >> Hmm, pride is heavy, boy.

[00:23:41]A true strong forklift knows when to lift others instead of competing with them. Teamwork lifts everyone higher.

[00:23:47]Bolt realizing the feud is hurting Yak's kids. Bolt's right there. Used to think my beating Yak was the win, but my rivalry makes things harder for them, too. Bolt decides to reach out. Yak, can we talk like adult forklifts? Yeah, I think so. They put the feud behind them.

[00:24:02]I was wrong, you're not the enemy. Me, too. I let pride steer me. We're stronger working together. They team up and get things done. Teamwork lifts everyone. Two forklifts, one warehouse, one team. All is well at home and at work. Okay, kids are happy. Happy kids, smooth shifts, that's the best payload.

[00:24:18]Friends, partners for life. Beep.

[00:24:20]They're high-fiving with the forklift things. Rivals no more, team forever.

[00:24:26]I mean there's really only one final final way to take this.

[00:24:32]So, I've said finally in a thought bubble, this has all been a weird daydream in the person on the right of this image from the um live stream image release. So, let's Let's see what it does with this. Do you like this personality? I kind of do, but I don't want to press anything because I like to let fate decide.

[00:24:49]This video is going to be way too long.

[00:24:51]I I do believe I said click off though like a few tests ago. So, if you're still watching, that's on you. It's not It's not my fault.

[00:24:58]>> [laughter] >> What the heck? Oh, what the What? No, it just added it added the it added the OpenAI image into the comic.

[00:25:08]What a weird daydream.

[00:25:10]Interesting. Okay, that was not what I expected, but I'll take it because it was creative.

[00:25:18]Wow. Regardless, this is going to conclude our first look and test of the new GPT images 2.0. I would say we did a few interesting things. What really stuck out to me the most as like, oh wow, was the sprite sheet where we had 10 different cars, 10 views of each specific car, and the level of detail here that it did, and actually how usable this would be. I mean, you could throw this into code actually. You'd probably have to slice and dice them in terms of getting each of these saved as an individual item, but it's trivial to at this point basically generate a sprite sheet like this, put it into a coding agent, and then be like, okay, make me this simple low-poly game. And And the other views could be used for like scenery cars. Or if one passes you, you see it from this angle and then this angle, and then like the possibilities are vast for actually using this in more productive ways than just like funny stuff as well. But we do like the funny stuff. I would say the actual thumbnails that it did here are quite all right. Um the TLC poster. This was a little disturbing. Maybe, you know, uh this it did not understand which vision I was referring to. And of course we had the forklift comic which just came from a place that I don't even know, but it did come to me. And then um this which I'm kind of fighting myself to not just post on X. I may as it seems fitting for that platform. So, regardless, that was our first look and test of GPT images 2.0, a very creatively named model that is quite good. I may clickbait this with insane, which I feel okay about because image models don't come out as often, so there's less of them.

[00:26:53]Um meaning more insane allotment. All right, that's going to wrap it up. If you have any questions, please feel free to leave them in the comments unless they are specifically related to where I came up with this random forklift idea.

[00:27:04]And thanks for watching.