A sharp, attractive generated portrait can still be the wrong photograph. The face may be plausible without being the person who supplied the reference photos.
I build CMeIn, an app for generating photos of yourself. Our upload flow checks normal photos against five poses and requires at least two poses before generation. That is an input constraint, not proof that the output preserves identity. The distinction matters when designing a review process for AI dating photos.
Here is a review rubric I would use before calling an output suitable. It is a proposed checklist, not a benchmark result or a claim that we have automated these checks.
Separate four questions
- Is the image technically usable? Check focus, lighting, hands and obvious visual artifacts.
- Does the face match the references? Compare face shape, hairline, nose, jaw and stable distinguishing features. Do not substitute attractiveness for similarity.
- Does the wider body still represent the person? A matching face does not excuse an invented build or a changed apparent age.
- Does the scene imply something untrue? A convincing mountain photo can still suggest an activity the person never does.
Those questions should not collapse into one score. A high score for image quality cannot compensate for a failed likeness check.
An illustrative review record
Imagine a generated cafe image with convincing light and a clean background, but a narrower jaw than either reference. The useful record is not "good image, 8/10." It is "technical quality: usable; likeness: reject; reason: jaw shape differs." That gives the reviewer a decision they can explain without pretending that a subjective rating is a scientific measure.
This example is hypothetical. I am not describing a customer or reporting a measured error rate.
Compare the same things
Put the reference and output side by side at similar face sizes. Compare more than one reference angle where possible. If the expression or camera angle makes a feature hard to judge, mark it uncertain rather than assuming success. A person who knows the subject can help, but their opinion is still a review, not a formal identity guarantee.
Do not measure success only by whether the image looks generated. "Could pass as a photograph" and "looks like the person" are different acceptance criteria.
Keep the failure useful
A rejection should say what needs attention: another source angle, clearer input, a different output, or using a real photograph instead. Adding more versions of the same front-facing input may not solve a missing-view problem.
The product context for this rubric is here: https://cmein.com/ai-dating-photos
The practical takeaway is simple: design the review so a beautiful wrong face can fail. Otherwise the process rewards polish while quietly dropping the requirement the user cared about.
This article was drafted with AI assistance.
