COMPANION MATERIALS · 23 OCTOBER 2026

Learning to See Beauty

Personalized Ranking from Portrait Comparisons

Can preferences learned from one observer’s portrait comparisons transfer to new photographs and an unseen person?

40observed choices

52source photographs

13identities

7methods

01 / FINDINGS

More stable scores did not reliably predict preference.

Models were fitted to 24 choices and evaluated on 12 comparisons using different photographs. Excluding one held-out tie, the main model matched 3/11 choices. Removing only the consistency term from the same representation yielded 2/11; an identity-only baseline matched 9/11.

Tongdak Cheonsa · held out from fitting

The observer preferred Tongdak Cheonsa to all four scheduled opponents. The main model missed all four comparisons and ranked her 12th among the 13 panels. These four choices do not establish the observer’s complete ranking of all 13 people.

The consistency term reduced within-identity score variance from 0.3254 to 0.000472 on fitting photographs, and from 0.6704 to 0.4036 on held-out photographs. Rankings also changed substantially when refitting on the original framing or changing the encoder. The rankings below are model outputs.

Study design with 24 fitting comparisons, 12 held-out comparisons, and four repeated comparisons — Figure 1 · The completed study design. Tongdak Cheonsa’s photographs and responses were excluded from personalization.

02 / MODEL PREDICTIONS

Same portraits, different models

Panels are ordered by the mean score of two held-out photographs. Original-framing conditions are compared with the same responses collected using cropped photographs; they are not a separate validation of preferences for the original images.

Encoder Image framing Method

Model rank / held-out panelMean of two scores

Loading records…

Sources and licenses for all 52 photographs

Crops are derivatives of the listed source images. Actress photographs retain their individual Creative Commons licenses. No Creative Commons license is assigned to the user-supplied Tongdak Cheonsa photographs. Source dates may refer to publication rather than capture.

03 / FIGURES

Results and computations

Figures were generated from the saved features, predictions, and responses. Seven conceptual diagrams are editable in diagrams.net; quantitative plots can be regenerated with the included Python code.

Tongdak Cheonsa and her four scheduled opponents, with predicted probabilities under each framing condition — Figure 4 · Four observed target preferences and the model predictions. Changing framing with fixed weights and refitting the model are separate conditions.

Agreement of seven models with all 12 held-out choices — Figure 3 · Every held-out response. The tied response is excluded from the accuracy denominator.

Within-identity differences in photograph scores on fitting and held-out images — Figure 6 · Stability on fitting photographs does not transfer equally to new photographs.

Coefficient-plane slice of the actual objective with consistency and preference-fit contours — Figure 9 · A slice of the actual objective. The line connects two saved solutions; it is not an optimization trajectory.

Download seven editable diagrams ↗

04 / OBSERVED CHOICES

All 40 observed choices

The session contained 24 fitting comparisons, 12 comparisons with different photographs, and four repeats with left and right reversed. Names and split labels were hidden from the observer. All four repeated choices agreed with their originals.

Show responses

Order	Split	Left	Right	Choice

05 / REPRODUCIBILITY

Paper, source images, and code

All archives extract into one chicken-angel-beauty folder. Model weights are not bundled; rerunning feature extraction downloads the pinned public checkpoints. The saved features support reproducing fitting, evaluation, and figures.

Paper PDF Reproduction guide Frozen protocol Full evaluation JSON Observed responses CSV Source credits CSV

Each archive includes a SHA-256 manifest. Downloaded literature PDFs are not redistributed. This is a case study of one observer and a fixed photograph collection, presented as an independent preprint using the official NeurIPS 2026 format.