Every AI photo tool produces the occasional below-bar image: a smoothed print, a warped hand, a garment rendered at the wrong length. The difference between tools is what happens next. Some ship it and let you burn a credit finding out. Some route it to a human retouch queue measured in business days. Quality mode does neither: it catches the defect in seconds and re-shoots, automatically.
Review one: the measurable things
Objective, instant checks run first — color similarity between your source garment and the output, a resolution floor, aspect-ratio sanity, and exposure. These are cheap and unambiguous, and they catch a class of failure no opinion is needed for.
Review two: an AI judge with your original photo in hand
Then a vision model sees your original reference next to the generated output and answers four strict questions:
- Garment match — same category, colour, pattern and construction: neckline, sleeves, slits, and length judged from your photo's own proportions. A thigh-length kurti rendered knee-length fails, even if everything else is perfect.
- Details preserved — prints, embroidery and borders sharp and faithful, not smoothed or invented.
- Person OK — natural face, correct anatomy, realistic adult proportions.
- Image clean — no artifacts, text, watermarks or duplicated limbs.
Accessories get their own reviewer with presentation criteria — a handbag is carried, not worn, and judging it like clothing would fail correct shots.
The retry is steered, not a re-roll
When the review fails, the judge's named issue and the expected construction feed the second attempt — the re-shoot fixes the identified problem rather than rolling the dice again. Those retries are absorbed by CatalogX (Pro gets 3 free retries a day, Business 5), because a below-bar image should be our cost, not yours.
Fail-safe by design
If the reviewer is ever unavailable, that is never treated as a pass. And every verdict is stored with the generation, so quality is auditable over time rather than anecdotal.