Because the model has defaults that your prompt does not override, and those defaults are consistent rather than random. That distinction decides whether retrying is worth anything.
We specified a tight framing: cropped from eyebrows to cheekbones, subject looking at camera. Zero of five renders met it. Every single one came back cropped forehead to chin, with the subject not looking at camera. That included the verification render taken after a fix, which is the detail that matters.
If those failures had been random variance, generating more images would eventually produce correct ones, and at four cents each you would simply buy your way through it. They were not random. They were systematic provider behavior, the model's own idea of what a portrait is, reasserting itself against the instruction. Retries against a systematic bias just buy you more of the same picture.
The fixes are structural, not financial. Either rewrite the prompt so the framing is expressed in terms the model actually responds to, or accept the model's framing and add a deterministic post processing crop that produces your spec from its output. The second is usually more reliable, because it does not depend on winning an argument with a model about what a portrait is.
Budget accordingly. Price the renders at four cents and price the pipeline work at whatever an engineer costs, because that is the line item that decides whether you ship.