H-ARN: Why a Beauty Rating Needs More Than an Average
Can two photos receive the same average rating while people disagree very differently about them?
One average can hide different opinions
Imagine a panel whose ratings cluster near the middle and another panel split between lower and higher ratings. The means can be similar even though their disagreement is different. A distribution-aware product could describe that distinction, provided its spread estimate has actually been evaluated against individual human ratings. A language model saying a face is ‘polarizing’ is not such an evaluation.
Disagreement and error answer different questions
Rater spread describes differences between people responding to a photograph. Model error concerns whether a prediction matches the intended target. A broad predicted distribution does not automatically mean the model is unreliable; a narrow one does not prove the mean is accurate. We therefore require separate distribution and error checks in our methodology.
What we would validate
Our proposed evaluation keeps individual ratings, evaluates the mean on held-out identities, and evaluates the predicted distribution with proper scoring rules. Any interval used to describe prediction error needs a coverage check of its own. We would disclose the rating question, scale, panel composition and exclusions so the reader knows whose perception is being modeled.
Current product status
We reference H-ARN as a design direction. We have not deployed its model, learned a rater-disagreement head or generated calibrated personal intervals. The paper-linked code address returned 404 at review. Current reports use six evidence-grounded observation dimensions; they do not label a person as universally attractive or assign a population consensus.
What we have adopted
Informs our proposed distribution evaluation. The paper-linked repository returned 404 on 30 October 2026; no H-ARN model is deployed here.
Our public methodology documents the deployed evidence layers and proposed validation protocol. For a concrete example, inspect the complete sample report. The example is illustrative, not an experimental validation result.
Continue reading
AI vs Human Attractiveness Ratings: Ranking Is Not Calibration
Why agreement in ranking does not mean an AI’s absolute attractiveness score matches human ratings, and how to read uncalibrated photo judgments.
Validating Face Analysis: Independent People, Independent Ratings
What an independent human-rating evaluation must establish before a face-analysis product claims validated predictive performance.