The QA filter — generate many, grade every one, ship few
Locks reduce drift; they don't eliminate it. Some fraction of every batch comes back with a rogue purple, a wrong crop, a watermark ghost, six fingers on the barista. The unprofessional responses are the two extremes: ship everything because the deadline says so, or eyeball fifty images at 6pm and trust your tired taste. The studio response is a funnel with a gate at each stage:
Generate 50 → machine filter to ~12 → human pick 3–5 → ship.
At batch-workhorse prices — cheap enough per image, as of mid-2026, that a fifty-candidate batch costs a few dollars, though verify current rates before you bill — overgenerating is cheap. What's expensive is reviewing — so the machine pass exists to spend cheap compute protecting your expensive attention.
The machine pass: objective checks only
A script can't judge whether an image is good. It can judge, instantly and identically for every asset, whether an image is disqualified:
- Palette distance — is each asset's dominant color actually near a kit color, or did the model wander off-gamut? Hex math, next two drills.
- Crop spec — is the aspect ratio the one the deliverable needs? A 16:9 render is not a 4:5 feed post, however pretty.
- Artifact flags — watermark ghosts, garbled text regions, the standard model tells.
Each check is pass/fail with a named reason. "CUT: off-palette (d=81)" is a fact you can defend; "I didn't like it" is not.
The human pass: judgment, on the survivors
The dozen survivors get the pass only you can run — composition, storytelling, whether the barista's smile reads warm or unhinged. This is the pass that most AI-design training skips entirely: the market teaches generating; almost nobody teaches grading — critique, selection, and direction of model output as a discipline. Which is odd, because directing is the actual job. The model is the fastest, most literal-minded production artist you have ever art-directed, and an art director who approves everything is not directing.
Write the reason down for the picks too. "Chose C05: strongest negative space for headline, mascot pose matches campaign line." When the client asks why these five, the answer is a paragraph, not a shrug — and when fifty become five hundred next quarter, your QA filter is the only reason the job still fits in a week.
Next: defend a pick in writing, then repair a QA filter with a flipped threshold that was shipping the drift and cutting the heroes.