Draft to Render

Most writing about image models covers the output. These notes are about the input — the sketch, the half-formed sentence, the settings nobody picked deliberately — and what a model does with each.

The three pieces here came out of building the input side of a GPT Image 2.5 interface: a sketch mode, a prompt rewriter that also proposes settings, and a generation-over-generation comparison that had to be measured because the release notes do not answer it.

If you want the running version rather than the notes, there is a browser generator that runs the first image with no account, and the longer write-ups live on its blog.

Notes

• A sketch is structure, not content Why handwriting inside a drawing has to be read as instruction rather than reproduced, and what that costs to get wrong.
• A prompt rewriter that proposes settings too Rewriting the words is the easy half. Guessing the aspect ratio and the quality tier is where it earns its place.
• What actually changed between generations Release notes announce; they do not measure. Three differences worth knowing and one that is mostly marketing.