
Ten edits in a row, four image models
One model kept its picture together over ten edits and three did not. On every measure of how far the frame had moved from its own seed, FLUX.3 Image ended closest to it.
Benchmarks
We run public models on real architectural tasks and publish what we measure. Every run carries its date, the exact version and endpoint of everything tested, and the prompts needed to repeat it.

One model kept its picture together over ten edits and three did not. On every measure of how far the frame had moved from its own seed, FLUX.3 Image ended closest to it.

Gemini 3.1 Flash Lite Image is about three times faster than the other two: a median of 5.0 seconds against 15.2 and 15.8. It returns the same pixel count as Gemini 3.1 Flash Image.

No model repeats itself. Composition overlap between two runs of the same brief averages 0.07 out of 1.0: you do not get the same picture back with small differences, you get a different picture. If a frame is worth keeping, keep it; you cannot re-roll your way back to it.

There is no winner. The two approaches fail in opposite directions, and which failure you can live with depends on what you are doing.
Models change under their own names. Every run states when it was executed and against which endpoint and parameters.
Seed images, prompts verbatim, masks, the harness and the measurement script are published. If you cannot repeat it, it is not a benchmark.
Models are good at different things. We publish the numbers, including where a model loses, and leave the ranking to the job you have.
The eight briefs of the model comparison put through Meta Muse Image 1.0, measured by the same scripts so the published July numbers stay comparable rather than being rewritten.
Run not yet executedWhether an exterior, its floor plan and its interiors read as the same building when generated from one brief.
Not started