Published: 2026-09-09

Sketch-to-layout in ChatGPT Image 2.5, and a real test of edit consistency

Name: Sketch-to-layout in ChatGPT Image 2.5, and a real test of edit consistency
Uploaded: 2026-09-09
Description: A hands-on look at the sketch feature in ChatGPT's updated image model — drawing a rough room layout with coloured shapes, describing what each shape means, and getting an aerial 3D render — plus a deliberate test of how much detail survives repeated edits.

Chapters / key moments (click to jump — plays here on the page)

The technique worth copying is the pairing: draw the spatial relationships, then describe what each shape means. Coloured rectangles and circles on a canvas carry position and scale, which is exactly what prose is bad at; the accompanying text says which blob is the desk and which are the cameras. The output is a labelled aerial 3D layout. The other half is a genuine test rather than a demo — the reviewer makes a series of small edits specifically to measure drift, because his standing complaint about the previous model is that a face becomes unrecognisable by the third or fourth generation.

Source video

"ChatGPT Image 2.5 Just Dropped — Here's Everything That's New" by Bart Slodyczka — Watch on YouTube →

Gotchas & Caveats

This is an image-generation feature rather than agent tooling. It earns a page here for the sketch-plus-description technique and the edit-consistency measurement, both of which generalise to any spatial or layout task you hand a model.
Availability was still rolling out when recorded; if you do not see the updated images tab, you do not have it yet.
The consistency finding comes from one sequence of edits on one image. It is a real observation and a small sample.
Generation timings were measured on the reviewer's account under normal load and will vary.

Key Takeaways

The sketch feature lives behind the plus icon in the images tab, opening a canvas with shape and freehand tools. Access was still rolling out at the time of recording — the updated images-tab icon is the tell.
Shapes carry spatial intent, text carries meaning. The working prompt names each shape by colour and says what it represents: white square is floor space, green rectangle is the desk with two monitors, yellow circles are camera positions, purple rectangle is where you sit.
Generation times, measured: roughly a minute and a half for the first render, about 1:50 for a subsequent edit. Progress climbs at roughly 1% per second, so the bar is a usable estimate.
Consistency across edits held. Flicking between generations, the reviewer finds essentially everything preserved between the original and the edited version — the only visible difference being lighting. That directly addresses the complained-about failure in GPT Image 2, where a subject degraded by the third or fourth generation.
It is not perfect and he says so: the generated signage came out sideways, and some details in the render are inventions rather than reproductions of what was drawn.
Edits are targeted by region. Marking an area and describing the change ("in that red circle, put a really tall plant") is how small changes are requested, rather than re-prompting the whole image.

Sketch-to-layout in ChatGPT Image 2.5, and a real test of edit consistency

Gotchas & Caveats

Key Takeaways

More ChatGPT news

Go deeper: ChatGPT guides