Face Symmetry Test: What Changed When We Changed the File?

On 26 October 2026, we uploaded five versions of one portrait to our face symmetry test, three times each. The unchanged file scored 98.0 on all three uploads. A darker tone curve scored 97.6, a rotated image 98.1, a smaller image 98.1, and a JPEG copy 97.5. Those are readings from edited files, not measurements of a changing face.

The setup

This was an image-processing experiment using an existing public portrait, not a participant reshoot. We used the same loaded production page in Chrome throughout. Each variant was generated directly from the unchanged RGB image; we did not apply edits cumulatively. The baseline was a lossless PNG at 512 by 512 pixels.

The darker version used the channel mapping round(255 x(value/255)^1.4). Rotation used Pillow BICUBIC resampling on a fixed canvas with white fill. The smaller image used LANCZOS resampling. The JPEG copy used encoder quality 20 with chroma subsampling 2 and no optimization. These settings describe our prepared files; quality 20 is not a percentage of facial symmetry or a measured percentage of data lost.

Synthetic darkening is not a test of physical illumination. Rotating a flat image is not turning a head in three dimensions; that condition also includes interpolation and canvas fill. We did not isolate those components.

Every upload, including all repeats

All observations below were made on 26 October 2026, Beijing time. Scores are copied at the precision displayed by the tool.

Ten unchanged-file uploads before rejudging the differences

On 26 October 2026, we opened a fresh Chrome tab and uploaded the exact unchanged baseline file ten consecutive times. We did not reload the page or alter the file between these ten uploads. All ten completed. This phase used a new tab because the earlier experiment tab had closed.

For each field separately, our observed repeat range was its maximum minus its minimum across these ten displayed readings. Overall: 98.0-98.0=0.0; eyes: 98.0-98.0=0.0; brows: 98.2-98.2=0.0; nose: 98.2-98.2=0.0; mouth: 97.1-97.1=0.0; jaw: 98.6-98.6=0.0.

Our stated gate is absolute difference greater than the corresponding repeat range. It applies independently to each field. All four total-score differences pass this gate:darkening 0.4>0.0,rotation 0.1>0.0,resampling 0.1>0.0,and JPEG 0.5>0.0. Every feature difference in the comparison below also exceeds its own observed 0.0 range.

Passing this gate means a difference exceeds the variation observed in these ten displayed outputs. It does not establish statistical significance, zero detector noise, a confidence interval, anatomical accuracy or a typical error for other images and sessions. Rounded outputs may hide smaller coordinate changes.

What changed relative to the baseline

On 26 October 2026, the gamma 1.4 file changed the displayed total by - 0.4 points: 97.6 minus 98.0. The mouth changed by - 0.9 points: 96.2 minus 97.1. The eye reading moved in the other direction, from 98.0 to 98.1.

On the same date, the + 10-degree image-plane rotation changed the total by + 0.1 points: 98.1 minus 98.0. That small total movement accompanied an eye increase of 1.1 points and a mouth decrease of 0.9 points. The total did not describe every feature moving together.

The 256-pixel image also returned 98.1 on 26 October 2026. Its nose reading rose by 0.7 points, from 98.2 to 98.9, while the jaw fell by 0.9 points, from 98.6 to 97.7. A matching total across two edited inputs did not mean matching feature readings.

The quality 20 JPEG returned 97.5 on 26 October 2026, a change of - 0.5 points from 98.0. Its jaw reading fell by 1.7 points, from 98.6 to 96.9, even as the eye reading rose by 0.6 points. These are displayed-score differences, not anatomical measurements.

All 15 attempts completed; no failed attempt was omitted. Within each condition, all three displayed six-field vectors matched. That observation concerns the displayed precision in this session, not equality of every underlying landmark coordinate.

Each nonzero difference above is retained only as exceeding its corresponding observed 0.0-point repeat range from ten uploads. No zero-difference claim is presented as a signal.

What this record does not establish

There was no physical lighting experiment, head-turn experiment, reference anatomical measurement or human attractiveness panel. The record does not establish a typical error, a passing score, a diagnosis or an effective intervention. We did not compare the underlying landmark coordinates directly. We changed files and observed outputs; we did not determine the internal cause of each shift.

For this comparison, keep the condition name attached to the entire six-field result. Choosing only the highest total would discard the feature changes that make these particular runs informative. This is photo-based geometric scoring, not a beauty rating or medical assessment.

Replay record

The input manifest records dimensions, processing settings and file hashes. The timestamped run log records every upload and its status. No outside measurement dataset supplied the scores in this article.

All original fifteen observations · Ten-repeat record and field-specific comparisons · Prepared input settings and hashes

Open the face symmetry test · Compare the other photo variables