ChatGPT Images 2.5 Is Here. Together With Astra It’s Crazy

ChatGPT Image 2.5 is here. And honestly, some of the results are pretty damn good. But the part that gets really crazy is what happens when you combine it with GPT-6 Astra. Because up until now, most AI gives you advice or it does a one-off task, right? Astra can now pilot your computer. So I'm going to show you the new ChatGPT Image feature, which is amazing. But then we're going to take one real job across 5 different apps. And what started with ChatGPT image is gonna turn into this, a full poster that has been submitted for print with our online printing provider. But I want you to know this, this is not really a video about making a poster. It's a video about one job surviving every handoff between apps and ending as something actually usable. Because we basically started with 2 ordinary photos, used Images 2.5 to create and edit the visuals. Then Astra took it over to Canva, created the whole poster, exported it for print, uploaded it to the online printing site. And heck, while we were at it, we even had it handle the screen recording of the entire process. All of that just worked. The full run took about half an hour and needed one permission click. So first, let me show you what Images 2.5 can do on its own. And then what happens when Astra takes over. Alright, so welcome to another Week in AI. OpenAI released Images 2.5. We got Astra last week. And well, they basically said that this Images model brings sharper details, more natural lighting and textures. Better reference image handling. That's a big one because you can make images of yourself or your friends and they actually look like them. And more precise editing. And finally, you can make multiple changes to the same image and it actually maintains that history. This was the most annoying thing with the previous image model that was really good at editing. But when you edited 2 things, it forgot about the first edit and you went back to square one every time you wanted to change something. So additionally to that, they claim that this is about 50% faster than the previous model. Feels faster. I don't know if it's exactly 50%. But the most important question here is simple. What can you actually make with this that would be interesting to you? So let's start with some examples of what we made. I mean, we just started off with the YouTube thing because it's top of mind as we're making a YouTube video. One example that immediately got me is we went from this, you know, very serious and sexy image right here to these thumbnails. And that actually looks like me, which I gotta say, the fact that it can preserve identity more than any image model before It's huge because you can now do images of yourself and your friends. So if I just Google a person, and I take an image of that person, it does it. And wouldn't you agree that this is the same person? Whereas, you know how it went with previous models, it was slightly different. The difference in identity has never been smaller, though. It's really good at editing, as you can see. So what we did is we took an object and we edited that with multiple prompts. And look at that, we started with this naked mug. And we just wanted to change one thing and leave everything else alone. That's the color. Make the mug forest green. Look at that, did it. Even if I move between these images, can you see how it maintains, well, almost the perfectly same position in the frame? That used to not be the case. But then while we said keep everything else the same, look at that. The lighting shifted, the structure of the table shifted. So it's not perfect. I'm not sitting here saying that, you know, it's gonna keep everything the way it is. I'm just saying if you want consistency, it's getting better at it. So next up, we just created a simple poster and we're like, oh, instead of a coffee, hey, how about a croissant? Boom, perfect change. See how everything else is consistent? So this is new and very nice. Wow, it's very nice, for the lack of a better word. Really very nice. That's how deep my vocabulary goes. It's an opulent iteration on the technological state-of-the-art image generation and editing capabilities. That makes sense. I don't know. I still prefer nice. Nice. Okay, but then on to what I think is really cool. It has a new feature where you can sketch things out really simply, and it turns them into reality. Okay, so look, this would be the example. You kind of make these boxes and boom, you have a poster for your pottery class. Let's try this ourselves right here. This one is really fun. All you need to do is go into ChatGPT, be in either chat or work. And with@, you're going to say sketch. With this new sketching functionality, look at that, I could draw the one thing that I know how to draw. Ish. I mean, very ish. Let me know if you can identify this character. Hell yeah. What's this? It's an art project. Okay, I like it. Picasso. That's Homer Simpson. Hey, golden star to you if you managed to actually identify it by this drawing. Make this into a realistic person with yellow-toned skin. Make this into a person with yellow skin. And based off our sketch, we can make things. Okay, as that generates, I also want to show you these editing features, because if I go on here, I could mark things up now. So if I was like, hey, this guy absolutely needs a cylinder, and then we need text across it saying good day, just make those edits, send the message. Let's see what we get here. Oh, and it asks for clarity, because as if this wasn't obviously a red top hat and a good day caption on top of the headshot. I mean, hey, have you never done this before, chat? Come on. And there it is. He has a hat and it says good day. But see, this is exactly how it works. And now if I wanted to say, okay, make the hat burgundy and add a cigar, I want you to pay attention to what he looks like and how this flows. We went from this image to corporate headshots to whatever this is. And now we're making edits. It's a continuous flow. If you ever tried this before in chat, you would have experienced that, hey, okay, first edit it makes and it's pretty good. Sure, there's improvements, but the main thing is If you tried to continue the workflow, whether through language or sketching, which is new, this image would have abandoned everything that happened here along the way. So this corporate variation. So maybe you would have seen him in the brown sweater again, with the top hat, and you'd be like, what was going on? As it wasn't carrying forward the context between the images. Now it does that. Yay! Can we have a little like confetti thing? Yay! There he is, all the edits in place. Still looking like the same person, staying true to his original identity. Let's have a look at back at the ooh, the yellow person that was apparently it may apparently violate their guardrails concerning similarity to first party content. This looks too much like Homer Simpson. That's what I'm sensing here from their guidelines. All right, so I did too good of a job of drawing him. Let's see what we get anyway. I wanted to check for this because previously they were very strict on their guardrails. Let's see. see where we land here. Usually on launch, it's a bit more loose, and then they tighten it up over time. Come on, 99%. You can do it. Come on, chat. You can do it. Do it. Just do it. Well, it's refusing to do it still. So I tried regenerating a few times. So let's just try again. Add sketch. Here's my second character. Now, this is no particular person. It's just sort of, I don't know, Mr. Worm. And he too has a red cylinder inspired by the good old Worms game. Anybody remember this one? This would be such a fun multiplayer thing. Guys, I apologize for these examples today. I swear we have some good stuff. We have a proper workflow in the second video. I just felt like trying random stuff here. So I did. Oh my God, is he made out of tooth? Okay, so my hope is you're going to use this for something proper. But yeah, editing images, amazing tool for that now. Okay. And hey, if you're enjoying practical and sometimes lighthearted videos like this, make sure to subscribe to the channel. I do this every week. Now additionally, if you want to find some fun things to do, they have this template gallery. Just go to images here on the left. And by the way, if you see this updated badge, that means that, well, you have the new 2.5 version. It's also available for the API. I already connected it to my agent, can create things in bulk if I want. But there you go. There's a bunch of things like this is really the answer to, hey, what useful things could I create? Well, here's a bunch of ideas. So all of this makes image generation more useful. And now this is clearly the state-of-the-art image generator that I would recommend to 10 out of 10 people really, except maybe if you're looking for the artistic look of Midjourney. It's an option. But the next question is really whether AI can take the image and continue the job because oftentimes creating something is just the first step. Okay. And we'll move away from these fun examples here in the beginning. But usually creating and editing an image is just a part of a workflow, right? Then you're going to do something with the image. That's the whole point, whether it's a headshot right here that you're going to upload straight to LinkedIn, or you're going to continue the workflow. So we tested a bunch of things. And what we found is that once you hand the keys to your computer to Astra to remote control it concretely through the ChatGPT desktop app, that you'll update to the newest version. And then when you switch over to the work tab, and then in the model switcher, you switch over to Astra. And by the way, I would recommend the medium setting. That's really the sweet spot between performance and cost. Many people have been talking about how high or extra high just gets you a few percentages more in certain benchmarks and use cases. And medium is really the best bang for the buck. And then here's what we did in our testing. We ran a bunch of different prompts. And this one impressed me the most because it's like, using 5 different apps in the process, and it just works. So we gave it one detailed brief on the workflow. We'll put that on screen right now. And we'll also share this prompt with you for you to try this yourself. Additionally, we gave it access to the tools it needed. Okay, and a clear rule for when to stop, what the result should look like. And yes, we were logged into Canva and to Dropbox along the way for this to work in the browser that this piloted. And it just moved through the job with 5 different apps. So what exactly does this workflow consist of? And what can you model? Well, we just started out by giving Astra access to the photos folder in Dropbox, literally just 2 images in Dropbox. And in the prompt, we described the poster, what we want as the output. As long as you're specific about the result you want, or in other words, what good looks like to you, then it will be able to find the routes to get to that result with any tools it might have or any tabs that you provide to it to work within. So then it took it into ChatGPT in the web, and it created this first image for the poster. Okay, so the brief was create a kitesurfing private lessons with Igor type of poster to promote it and find the online service that can print it and upload the image so it's ready for the order to be submitted. So as you can see, it first made the image, then it also edited the image, as I just showed you with the new Images 2.5, which is really impressive. Then it downloaded the image, and it wasn't done yet. It was like, okay, this is a good starting point. But usually graphic designers take this stuff and they use something like Photoshop or Canva. So in our case, it just moved over to Canva and did all the work with laying out the poster. I mean, check this out. Pretty sweet. Again, reminder, this was all one prompt and a permission and just hitting enter. All of this text stays editable so you can change it afterwards. That's the difference between a nice image and a useful working file. Okay. Now it exported the whole thing as a print-ready PDF and found a site that can print these sorts of things and uploaded the poster to their service. Now we have an 18 by 24 inch poster ready to continue through checkout. The whole thing took about half an hour including waiting and correction. We have not printed this, but it was like 2 clicks away. Heck, fun fact, during the production of this, we even asked it to stop the screen recording once it's done and it opened up OBS and clicked stop on that thing once the whole workflow was done. It's crazy. And it all just worked. And we didn't need like multiple attempts to get it done. Kind of put these things in and it's just reliable enough for you to hand these things off. That's new. And think about it. If previously you were like, okay, I'll make an image with AI and then I have to do something with it, then great, then automate it like that image making piece. But now it can do that and it can use Canva and then do the other thing that you wanted to do with that result while maintaining editability. Like you still have the ChatGPT chat where you could change things. You still have the Canva file where you could make changes. You can create these workflows that are non-destructive, because whatever happens along the way, you could revert each step or make changes, which makes working with AI so much simpler than having to blindly rely on some workflow that it just does in the background. It's like a legit coworker that you can hire. And all of that from one prompt? Pretty impressive. And look, this poster is just one example. The reusable idea is a job that has to move through several tools before it's actually finished. We're already trying these sister use cases like video production, because really, it can research, it can create the brief, it can script, it can generate the voice, It can edit the video, it can render it, it could do final quality control. You could still make the creative decisions. But you know, the rough work is done. And you would be there to judge the outputs and fine-tune it. Another example would be a weekly operating review where you pull the status or information from one application, you pull the financial numbers from another, and you pull the decisions that have been made from your own meeting notes. And you assemble everything into one review with priorities, owners, blockers, just a once a week report. Report on how things are progressing. It's really things that up until now you needed a project manager for. And now, wow, you can really piece these things together and have a working version. And the key here is that it can keep the same job intact from beginning to end. So is ChatGPT Images 2.5 good? Yeah, it can produce some excellent first images, gives you more ways to steer it. And it holds references and edits better than anything before. But this combination with Astra is one more interesting shift because instead of downloading the image and starting over in another tool, the image can become the first input in a larger job. So you can start thinking bigger because Astra can carry it across multiple apps, preserving the objective that you initially stated, and leave you with something you can actually continue using or something that gets you an actual result rather than AI helping with one step of a longer workflow. The one thing that I will say is that this was not fully autonomous. I needed to log into the apps first and give it permission once. To do all of this. But once you do that and figure these workflows out, you could schedule them to run regularly if that produces value to you. And there you go. That is pretty much everything I have for you today. Go create something amazing and let me know what you're doing with this combination of images and Astra. My name is Igor Pogany, and I hope you have a wonderful and productive day.