When two generated portraits look different, “better” is often too vague to explain the choice. One may have more convincing lighting while the other looks more like the person. A short rubric helps keep these judgments separate.
Disclosure: I maintain RetroPrompt. This note was prepared with AI assistance and proposes a manual review method; it is not a validated quality benchmark or a report of a user study.
Use four separate checks
Likeness: compare visible facial proportions, expression, glasses and other identifying details against the reference photo. Do not assume a polished face is an accurate face.
Style: ask whether the requested lighting, wardrobe and backdrop are recognizable. A stronger effect is not always a better result; the useful amount of styling depends on the purpose.
Artifacts: inspect hair boundaries, glasses, jewelry and any hands that appear. Also check background lines that pass near the subject. Review the exported image at a useful zoom level rather than only the small preview.
Usability: check the intended crop and display size. A convincing full-size portrait may lose its expression in a tiny circular avatar. The best creative image is not automatically the best profile image.
Prefer explicit decisions to a fake precision score
For each check, write “acceptable,” “needs revision,” or “reject,” followed by one concrete observation. For example: “needs revision: the glasses are narrower than in the source” is more useful than “likeness 8/10” when there is no shared scoring standard.
Keep the dimensions separate. Adding them into a single total can hide a serious likeness problem behind good lighting. If a result fails the use you care about most, do not let its other strengths erase that failure.
Make the comparison fair
Compare images at the same crop and display size, using the same reference photograph. Record which model or service produced each result when that information is available. If several variables changed, describe the comparison as exploratory instead of attributing the improvement to one prompt phrase.
Keep rejected outputs only as long as you need them and avoid sharing another person's reference photo without permission. Generated creative portraits should not be represented as historical photographs.
For a set of starting prompts, I maintain this collection: https://retroprompt.site/prompts/?utm_source=tilnote&utm_medium=referral . These are starting points for still-image experiments, not guarantees of likeness or evidence that every suggested prompt has been benchmarked.
A useful review note should tell you what to change next. If it cannot do that, make the observation more specific before generating more variants.
이 노트는 보도·비평·교육·연구 목적으로 작성되었습니다. 저작권 문의가 있으시면 에서 알려주세요.