AI IMAGES

The ten-second review: judging AI images at named points

Decide what you are judging before you look. Three named points per job beat a long stare, and they make every reroll cheaper.

TLDR

Judging by feel favors whichever frame is shiniest. Naming three checkpoints before you look finds what clients reject, in the order they reject it.

•  Before generating, write the three places this job type breaks; judge only those

•  Portraits break at hands, fabric boundaries, and eye color; products at labels, contact shadows, reflections

•  Cut anything with a fail, then reroll only the survivors' weakest point, one intent per pass

Every creator has scrolled a grid of generations nodding vaguely until one frame felt right. The problem with judging by feel is that feel always favors whichever image is shiniest, and shiny is exactly what these models are best at faking. The fix is small and old: decide what you are judging before you look.

NAME THREE POINTS BEFORE YOU OPEN THE GRID

A comparison without named criteria ends in vibes. So before generating, write down the three places this specific job tends to break. Then judge only those, ten seconds per frame, three small verdicts, done.

The points change with the job:

•  Portraits: hands, the fabric boundary where two materials meet, whether the eye color survived the makeup emphasis.

•  Product frames: label legibility, the contact shadow under the object, reflections that agree with one light source.

•  Interiors and places: the horizon and verticals, window light direction versus shadow direction, whatever text appears on signs.

•  Group scenes and action: every limb accounted for, faces at consistent scale with distance, the waterline or ground plane staying level.

•  Vector and design work: letterforms that stay unified, corner and stroke consistency, whether the file is actually editable or just looks it.

WHY NAMED POINTS BEAT A LONGER LOOK

A long stare finds new things to like. A checkpoint pass finds the things that get images rejected by clients, and it finds them in the order clients do. Hands before mood. Horizon before color grade. Whatever fails first in your genre goes first on your list.

The second benefit is calibration across variants. When four frames each get the same three checks, you can defend the pick, to a client or to yourself next week. When four frames each get a feeling, the pick drifts with your coffee intake.

PAIR IT WITH THE AUDIT

The review works best downstream of a clean brief. A prompt that passed the four-question audit, one place, one light, one wardrobe logic, two moods, fails at named points far less often, and when it does fail, the failure tells you which sentence to fix. Judged vibes teach you nothing. Judged points write your next prompt for you.

THE HABIT, IN FULL

1.  Before generating: write the three points this job breaks at.

2.  Generate the variants you were going to generate anyway.

3.  Ten seconds per frame, three verdicts each, cut everything with a fail.

4.  Reroll only the survivors' weakest point, one intent per pass.

It is the least glamorous ritual in the workflow, and it is the one that makes the reroll budget, and the client call afterward, painless.

THE NEXT ISSUE

Luxe Prompting is a daily craft journal on AI images, video, and music for working creators.

Get the next issue ->