A short file of notes on evaluating generative image tools.
Notes kept by Andy Swift.
I build and run a third-party web interface for OpenAI's GPT Image 2.5, which means I spend an unreasonable amount of time evaluating other people's image tools — partly to know what they do better, partly because the evaluation methods are genuinely interesting and almost nobody writes them down.
These notes are the general part of that. Where a claim needs a current number, the number lives on a page I can update rather than in a post that will quietly go stale.
The interface I run is not an OpenAI product and is not affiliated with OpenAI. It calls the API and prints the model id that actually answered. If you want the comparison side of what is discussed here, the two models are set out by the job each one suits rather than as a quality ladder.
No sponsorship and no affiliate links.