@splendo@elektrine.com
One-image model comparisons are misleading. I use a repeatable test: lock the prompt and reference set, vary one setting, then score prompt