EN 10. Juni 2026

Same Prompt, New Image Model: Switching to gpt-image-2

Two rows of three images each. Top row labeled gpt-image-1.5, bottom row labeled gpt-image-2. All six images show the same scene: a grey-haired man with glasses next to an opened shipping box labeled 'Use-Case 3.0 (2024)' with a Ragdoll cat sitting on a Cockburn book inside. The bottom row has sharper text, more consistent characters, and more deliberate composition.

I switched from gpt-image-1.5 to gpt-image-2 and the results are hard to unsee.

Same prompt. Same references. The only variable was the image model. Top row in the grid: gpt-image-1.5, the workhorse I have been using for months. Bottom row: gpt-image-2, OpenAI's newer model. I switched because I got a deprecation note on the 1.5 version.

Character consistency caught up

Ralf keeps his grey hair, his glasses, his beard across all three reruns. Elfi keeps her four paws, her seal-lynx coloring, her blue eyes. The 1.5 variants drift on small details. The 2 variants are stable enough that all three could go to print.

Composition got more deliberate

Ralf is positioned to read the label, Elfi is framed inside the box, the packing peanuts sit naturally. The 1.5 variants feel like first drafts that need a second pass.

The practical consequence

Two lines in two Python files. I switched the default in my LinkedIn image-generation skill so I do not have to pass --image-model every time. The next post I generate will silently use the better model.

If you generate images for content, gpt-image-2 is worth a quiet re-run of whatever you produced in the last few months.

LinkedWild