GPT-Image-2 dropped last week. I'll be the first to admit the previous models didn't quite hit the heights we were hoping for — but this one is different. Strong prompt adherence, great text, on par with Nano Banana Pro and Nano Banana 2. Where I think it pulls ahead is contextual understanding, especially when piecing architecture together.

Editing with prompts

Editing with prompts is something we've all become used to in Fenestra — change materials, swap the time of day, imagine a new form. GPT-Image-2 handles all of this well. But where it really excels is expanding on a single image. I'm exploring ideas for a Japanese-inspired teahouse, so I started here: "modern Japanese teahouse elevated off the ground in a dense bamboo forest, in the foreground a subtle Japanese garden , bright early morning, warm sunlight scene, exterior architecture photograph"

Competition boards from a single image

The result is great. Then it gets more interesting — I asked it to create what's essentially a competition board, and gave it free reign to fill out drawings, details, callouts, plans, elevations and sections.

"create a layout detailing this architecture with the main hero image at the top and under with plans, sections and elevations of the project with key call outs for specific details on the architecture" It's not perfect, but from a single image and one prompt, the result is incredible.

Multiple views in a single generation

Last thing I've been testing — can it creatively generate multiple views of a project from a single prompt in one generation? It does a fantastic job here, and it nails the text descriptions for each image too.

"create a triptych with three vertical views of this architecture, name the views accordingly, be creative around time of day, angles and details you're showing. I would like to see some detailed architectural drawings alongside aerials or alternative views in different weather conditions"

Try it out

This feels like just scratching the surface of what this model can do. Give it a go in Fenestra and let me know what you create.

Try GPT-Image-2→

Create now.