ChatGPT Image 2.5 Just Dropped (Full Breakdown)

summarized

TLDR

OpenAI released GPT Image 2.5 with multi-turn editing consistency and a sketch-to-image feature. In side-by-side tests, it produced more realistic and detailed images than Gemini's Nano Banana Pro, though generation speed was slower. The model is available via web, desktop app, and API with two variants: Flare (fast) and Sunburst (precision).

Key points

GPT Image 2.5 maintains image quality across multiple edits without degrading over time.

A new sketch feature lets users draw a rough sketch and have the model generate a refined image.

The model is available via API with two variants: Flare for speed and Sunburst for detailed creative work.

In a side-by-side comparison, GPT Image 2.5 produced more realistic thumbnails and infographics than Gemini's Nano Banana Pro.

The speaker noted that GPT Image 2.5 was slower than Nano Banana Pro despite claims of 50% speed improvement.

Tools mentioned

Techniques

  • multi-turn editing consistency
  • sketch-to-image generation
Transcript (captions)

0:00 OpenI just dropped ChatCBT image 2.5 along with a brand new way that we could generate images inside of Chhatchevt. So in this video I'm going to break down everything that is new with this. I'm

0:09 going to compare this to the previous Nano Banana model from Gemini as well as I'm going to test it out live in action so you could see just how capable it is. So without further ado, let's dive right

0:18 in. Now there are multiple ways that we can access the new ChateBT image model. First we could access it directly inside of the web browser right here. We could also access this directly inside of

0:28 chatbt desktop app as well. And then we could even use this inside of codeex which is also inside of the chatbt desktop app. And we could also access this via the API which we have two

0:39 different models. First we have GPT image 2.5 flare which brings the same quality editing and speed improvements while GPT image 2.5 Sunburst adds precision for more detailed creative

0:52 work with longer generation times. So, we're obviously going to be using the GBT 2.5 flare mode if we want a faster output, but we're going to use 2.5 sunburst for more complex images that

1:03 we're trying to generate. And again, this is available via the API. Now, let's go ahead and test this out. First, as you can see on the left hand side, we have this images section that we could

1:13 click on, and we're going to see this. And inside, we have multiple different templates along with trending templates of what other people are using in order to generate images. Of course, we could

1:23 always come to new chat and simply just tell it to generate an image. But if we want to deliberately come and generate assets, we could just come to this page right here. Now, one thing I want to

1:32 point out is we have this new feature added here. So, if I click this plus button and I scroll down, we're going to see this right here that says sketch. And we're able to literally draw with

1:42 our mouse or our finger in order to generate images. So, if I click on it, it's now going to pull up this canvas. We could either write some text, we could simply draw here. I'm just going

1:51 to try to draw a sunset. I am no artist at all, but I'm just going to come here. Okay, we got a sun. Draw some like sun rays here. Obviously, this looks absolutely terrible, but hopefully GPT

2:02 image 2.5 could take this kids drawing and turn it into like a Picasso painting or something. Let's just add like I don't know, we'll add some sky back here. I mean, this is actually comical.

2:12 Uh do not judge me based simply on my drawing skills. This is absolutely horrendous. And then now I'm simply going to add like some rolling hills or something like this to make it look like

2:20 it's a I don't know just like a field or something like that. All right. So I'm simply just going to send that off and let's see what Image Gen 2.5 generates for us. All right. And here is our

2:29 output. I'm genuinely surprised by what GPT image was able to come up with from my terrible sketch here. This is pretty cool. Obviously it's like cartoonish. But what I'm going to do is I'm simply

2:39 just going to say, "All right, this is really cool. I like the blueprint for this, but I want you to make this look very realistic instead of like this cartoon vibe." And here we go. This

2:47 looks obviously a lot more realistic. I think the sun I wish the sun was a bit bigger, but again, this does look more realistic with the rolling hills. I guess the sun is proportionate to kind

2:56 of the size that it should be here. So, this is really cool. And also, if I click on this edit button, this now pulls up this little interface. So, what I could do is I could simply come here,

3:05 I could click comment, um, select a specific part of the image I want to change, or I could erase the background, I could erase something, I could resize this, or I could just mark this up. And

3:14 let's say I want to select a specific area and I want to like remove the clouds right here or something. Let's just say I want you to remove the clouds in this circle. Now, one thing that's

3:24 pretty cool about Image Gen 2.5 is that we can edit images multiple times without it veering away from the original image that we gave it. So, if you've ever gone from a specific

3:36 reference image that you start with and make one or two changes, a lot of the times it could really start to look like a completely different image over time the more and more we edit it. And there

3:46 we go. We have the clouds right here removed. This is obviously nothing game-changing. This is just a new easy way to use chatbt image genen in order to make changes to the different images

3:56 that we make. So now let's go test a couple other use cases that are a little bit more complex so we could see the real power of GPT image 2.5. Now this is one of the most important parts of this

4:05 release. Multi-turn editing consistency because at this point image models like Chachabt Image 2, the previous model or Nano Banana Pro from Gemini, these models are already really good at

4:17 generating specific images and we already know that. But one thing that they really lack is consistency in the subjects as we make changes to these images because traditionally whenever we

4:28 regenerate an image a lot of the times after multiple generations it's going to veer away from the original reference image that we gave it. So according to this article it says during longer

4:38 chatbt conversations images 2.5 follow specific editing instructions more reliable across multiple edits. Earlier changes are more likely to stay consistent and each new edit builds on

4:50 the work you've already done without degrading quality over time. And this is the main point right here without degrading image quality over time. This is true for not only generating images

5:02 inside of the chatbt interface, but let's say we have a specific workflow inside of codeex or something where we have an agentic workflow generating a bunch of visual assets for us all at one

5:11 time. we need to be able to make multiple iterations on that without it veering away from our original reference. Now, this is an example that OpenAI gave us. So, GPT image 2 right

5:20 here, this is over 150 different generations. As you can see on the left hand side, the cube is kind of bouncing around. Whereas, GPT image 2.5 from 150 generated frames looks very, very smooth

5:32 and that is because this is consistent and it's matching the exact pose guides that we give it. And that is where the consistency really begins to be felt. Now, this is another very, very basic

5:41 example, but as you can see, these little like candles, they light up. These are all multiple generations, and nothing about this image changed at all aside from these different candles being

5:52 lit sequentially. Now, here's a specific use case that I want to show. Now, here's a very specific use case to me that I want to show. Basically, the main thing that I use AI image generation for

6:02 is simply my thumbnails. Every single one of these thumbnails that you see on my channel over the last like six to seven months has been generated by AI. And the image model that I've been

6:13 generating these with is Nano Banana Pro which I have plugged into Claude Code. I have a specific Claude skill created in order to recreate a similar thumbnail style and that is what I've been using

6:25 in order to generate these thumbnails. And I will say it's not very cheap with how many that I generate. Now, I wanted to test the two and simply just see how much better of a thumbnail I could make

6:34 with GPT Image 2.5 in order to test the capabilities. And here is an interesting output. So, now right here is the original thumbnail that Nano Banana Pro created for me. I think this is a really

6:45 good thumbnail, but it's nothing insane. Now, I think this is a really good thumbnail, but it feels slightly bland and slightly stale. So, I gave it to GPT image 2.5 with this very, very elaborate

6:57 prompt here saying, "Make this thumbnail better." And here is the output that I got. I think that this one personally looks much better. It's just much more pleasing to the eye. The colors pop a

7:08 bit more. Like this character right here, sun is shining off them a little bit more. I would 100% click on this thumbnail as opposed to this one. I just think that it looks much better and and

7:18 it did a much better job than Gemini Nano Banana Pro. Obviously, I don't think it's fair in order to compare the outputs when I simply just tell GPT Image 2.5 to make a thumbnail better. So

7:29 what I did is I gave each of these image models this exact same prompt. And then here is the outputs that it gave. First of all, this is what Gemini Nano Banana Pro gave us. It generated four of these

7:38 different images. I'm using Google Pix, which is a way for us to create and edit images with Nano Banana. And I think that these look obviously pretty solid, but you could definitely tell that this

7:47 is AI generated in terms of just like how this lady looks. Now to compare here is what GPT image 2.5 gave us. I think the skin texture just looks a lot better. Obviously, in my opinion, still

7:59 looks a little bit AI generated. Um, I think that this is very, very detailed, but let's test these side by side. Here is the sideby-side output. I'll let you be the judge, but I personally think

8:09 it's a no-brainer that Catch Image 2.5 on the left is much better than the ones that Nano Banana Pro generated on the right for us. Now, here is the output that we got. On the right hand side, we

8:20 obviously have Nano Banana Pro. I think that this is obviously really good. Looks much more AI generated in my opinion. Um, obviously like geographically like this is all correct.

8:28 It looks really good. Um, but this is something that we've been able to generate for quite some time already. However, on the left-hand side, we now have this infographic. I think this

8:37 looks much better. It followed my instructions quite a bit more. Um, basically the prompt that I gave, I said I want it to be a very realistic infographic. And this looks really good.

8:46 Like for example, this obviously you could tell it's slightly AI generated, but this looks like it's in much more realism style. has all of the information on the specific monuments,

8:55 which is exactly what we instructed. I think that this one looks really good. Honestly, I'm not going to lie to you. I think that this one I think that this one looks extremely solid. If I have to

9:03 give it a review out of 10, I'd give this probably like an 8 out of 10, whereas the one on the right is probably like a 6 out of 10. One thing that I'm noticing when I run these both side by

9:11 side at the exact same time is Nano Banana Pro is generating these faster. I don't know why, but GPT image 2.5 is much slower even though this should be 50% faster than the previous GBT image

9:23 model. So, just keep that in mind when you are using this. Anyways, that is the new Chat GBT image 2.5 model. I think this is a level up from the previous models. I don't think it's anything

9:32 game-changing. So, go ahead, test it out yourself. Let me know what you think in the comments and subscribe for more content just like this. See you in the next video. Cheers.

Frontier News · by Hyperjump Technology