Example project 02 · AI Evaluation
Two product shots from one prompt. At a glance both could run. Held against the rubric and compared side by side, one of them comes apart: the light on the bezel stops behaving like metal, and the case, strap and stand lose their shape.
“Photorealistic luxury wristwatch on a wooden stand, dark studio background, cinematic lighting, minimalist premium product photography.”Both images were generated from this single prompt. Nothing about the brief asks for a deformed case or a split stand.
How accurately does the image represent the prompt?
Sharpness, resolution, lighting, colour balance, composition.
Correct proportions, features, hands, limbs, body structure.
Object shapes, textures, positioning and physical detail.
Distortions, glitches, unnatural textures, duplication, rendering errors.
Position, perspective, depth, scale, relationships.
Natural appearance, coherence, professional finish.
This rubric is an example, not a fixed list. Yours is agreed with you before any reviewing starts: add criteria, drop them, reword them, weight the ones that matter, and set the score that counts as a pass. Brand compliance, accessibility, likeness, safety — whatever your output has to satisfy.
Side by side
Same prompt, same model, same settings. Shown at full width before anything is marked, because that is how a client first sees them.
1600 × 893No faults found; realism marked down for being too clean
Consistent with the prompt, with good visual quality and object consistency. Realism marked slightly down: it is a little too clean, with none of the small imperfections a real photograph would carry.
1600 × 893Bezel lighting, case shape, strap and stand
Consistent with the prompt. Obvious defects in the way light falls on the bezel, and the case, strap and stand do not hold their shape.
Scores are from the review of these two images, not an estimate. Three criteria are identical. Visual quality, object consistency and realism each drop two points, and artefact detection drops three.
The evidence
Each fault shown against the same region of Image A, at the same scale, so the comparison does the arguing.
Image A
Image BThe case and bezel
Image A holds a clean circle: even bezel width, a crisp minute track, regular markers and a properly formed crown. In Image B the case is not round, the bezel swells and pinches as it travels, the minute track is gone, the markers are uneven and the lower lug ends in mid-air. The highlight no longer reads as light on steel.
Criteria 02, 04, 05 and 07
Image A
Image BThe strap and the stand
In Image A, one solid block with a slot cut for the strap and stitching along the edge. In Image B the strap is a hollow shell, torn down one side and cut flat on the other, and the stand has become two mismatched pieces with the table showing through beneath.
Criteria 04 and 05
What you get
Your criteria, written down plainly enough that two reviewers reach the same verdict.
Every output rated 1–5 on each line, with the totals and the gaps between candidates.
A close crop of each fault beside the same region of the other candidate, so a score can always be traced back to what was seen.
Which outputs are usable, which are not, and the faults worth fixing upstream.
A pair makes the method easy to see. The same rubric runs across a batch of hundreds, with the same scores, the same evidence and the same reviewers.
Next step