Parametric Thinking

GPT Image 2.5 Sunburst vs. Flare

A few weeks ago, I compared some of the leading image generation models, including GPT 2.0, Nano Banana Pro, Meta Muse, and Seedream 5.0. Since then, GPT has released two new models, GPT 2.5 Sunburst and Flare, so I wanted to see how they compare. Both models cost exactly the same ($210.70 per 1,000 images), but they are quite different in terms of speed and what I would actually use them for. In my tests, Sunburst took around 1.5 to sometimes 2 times longer than Flare to generate an image.

For Flare, I think the biggest use case is speed. If you are building it into the backend of an app, quickly generating ideas, or going through a lot of iterations, Flare makes a lot of sense. I would also use it when I just want to test a concept without worrying too much about getting every detail right.

Sunburst feels more suited to final production work. It is slower, but I found it more precise, especially when the image/prompt has a lot of details that need to be followed correctly. It also works really well with text, so I can see it being useful for things like infographics (look at image 2 for reference).

Personally, I would choose Sunburst. I don't really need that much speed, but I do need more precision in my designs.

One thing I still don't like about GPT images is how contrasted and dark they can look. I hope this gets fixed in the next model. But they are definitely getting there. The images are becoming more realistic and slowly losing that recognizable "GPT aesthetic." You probably know what I mean. A lot of YouTube thumbnails right now have that look where you can immediately tell they were made with GPT.

For my own workflow, though, I would still stick with Nano Banana Pro. I can see myself using a hybrid workflow where I use GPT Image 2.5 Sunburst to get the precision and structure right, and then Nano Banana Pro to refine certain parts and get the final visual quality I want.

Have you tested Sunburst and Flare? What do you think?

FallAdventures
Parametric Animation (Rhino + Grasshopper) → Gen AI Design (ComfyUI + Nano Banana Pro + GPT Image 2.5)

Comment "Pattern" to learn the design process being this post

Testing some “State of the Art” image editing models. All the images were first outputs, which is why you see a lot more hallucinations than in my other posts. The base design is the last image from my previous post.

I keep doing these tests every 6 months to see if the new models are really that much better. For visualization, since GPT Image 2 which was released in April, there has been no major update (only talking about image generation for closed source models). So while Gen AI is getting better and faster every week, for visuals, things have not changed that much.

My thoughts on each of these frontier models:
- GPT Image 2: The consistency of the GPT model is the best, but for some reason, it always makes dark and crunchy images (closest word I could think), even after 5 months of the model being released.
- Nano Banana: For image quality, I still love Nano Banana Pro and Nano Banana 2. They hallucinate a lot, like you see in some of the images, so it’s tricky to get good images in one shot.
- Muse: The cheapest of them all is Muse by Meta, but it creates rubbish details. At first glance, the design looks the same, but it’s just slop. I would not use this model whatsoever.
- Seedream 5.0: It fits right in the middle of Muse and Nano Banana, but for its cost, it’s better to stick with Nano Banana 2.

In terms of cost, Nano Banana 2 is $67.00 / 1k images, GPT Image 2 is $211.00 / 1k images, Muse is $10.00 / 1k images, which is crazy cheap, and Seedream 5.0 is $90.00 / 1k images.

So overall, even after more than 6 months, I would still use Nano Banana 2 or Pro. I wish the images from GPT were not so AI looking. If they can solve that, I will move to GPT. What are your thoughts on these models?

There is a moment when you draw a curve and wonder what it will become. Will it be a surface, an extrusion, a wire cut, a reference line, or just a beautiful spline? The possibilities with curves are endless, and that’s what I love about curves. They are so fluid and non rigid and can become anything.

It’s like they say, a picture is worth 1,000 words. To me, a curve is worth 1,000 options. These are a few of my versions of how a curve can be translated into varied spaces.

Typewriter Pavilion
Inspired by the geometric arrangement of keys in a typewriter

Parametric Animation (Rhino + Grasshopper) → Gen AI Design (ComfyUI + Nano Banana Pro + GPT Image 2)

Comment “Pattern” to learn the design process being this post

My designs are influenced by my surroundings and what I see on a daily basis. The patterns, the textures, the composition, the symmetry - all of these catch my attention and help me in my design process.

Blendilion
Parametric Animation (Rhino + Grasshopper) → Gen AI Design (ComfyUI + Nano Banana Pro)

Comment "Pattern" to learn the design process being this post

A wall of work, and not one of these was prompted twice.

Comment “PATTERN” to know more about the workflow.

Different site, different brief, different material every time. Swipe to the end and you will see what sits underneath all of it: the same two part working file. Grasshopper builds the logic, ComfyUI renders it, and the design language stays mine in every single frame.

That is the whole argument for working this way. You are not hunting for a lucky image. You are building a system that answers to you, and once it exists it will give you a catalogue instead of a screenshot.