Notes on what generative image tools tell you, and what they leave you to find out.
Four short pieces about the parts of image generation that sit outside the model: what a free tier actually licenses you to ship, whether to run four images at once or one image four times, how to compare two models without fooling yourself, and how to tell a policy refusal from a thing the model simply cannot do.
None of it is about prompting. It is about the paperwork around the picture, which is where most of the wasted time actually goes.
A free image is not automatically a usable image. Three separate permissions have to line up before you can put one in front of a client.
Batching and iterating cost about the same and do completely different jobs. Picking the wrong one is the most common way to waste a budget.
Most model comparisons are decided before the first image is generated. A short method for running one that can actually change your mind.
A blocked generation and an impossible one look identical from the outside, and the fix for each is the opposite of the fix for the other.