Eve library

Appearance attributes, and what lighting does to them (2026-09-10)

Guide

Revision 1

On this page

Evidence for 182.C.19.02, which names eight: face shape, complexion, age presentation, hair, distinctive features, body proportions, wardrobe, and permitted changes.

Method#

The committed frame ../2026-09-10-openrouter-face-detection/two-foreground-subjects-plus-bystanders.png (1408×768, two foreground subjects and four bystanders), analysed with the models already staged in ~/.cache/oshun-human-video: YuNet for detection, face_landmarker.task for geometry, selfie_multiclass_256x256.tflite for the hair / face-skin / clothes segments, and pose_landmarker_lite.task for body proportions.

Then the same frame through five real lighting variants — two white-balance shifts (per-channel gain, which is what a colour temperature does to a camera) and two one-stop exposure changes — plus a ten-point sweep of smaller changes.

The finding: complexion is dominated by lighting#

Variant median lum red gain subject 0 L* / ITA subject 1 L* / ITA
reference 51.8 1.028 45.5 / −11.6 70.6 / +52.1
warm ≈3200 K 53.7 1.195 49.0 / −1.6 74.1 / +34.6
cool ≈7500 K 50.6 0.887 43.1 / −27.8 68.6 / +86.9
exposure +1 stop 94.9 1.017 82.7 / +46.6 not measurable
exposure −1 stop 25.9 1.028 22.4 / −66.5 36.9 / −55.6

The two subjects differ by 63.7° of Individual Typology Angle in the reference frame — two visibly different people.

A one-stop exposure change moves subject 0 by 58.2°. Almost the entire between-person gap, from nothing but the light. A modest white-balance shift moves it 10.0°.

So a complexion comparison between two frames lit differently is a lighting comparison wearing a complexion's name.

How close does the lighting have to be#

From the sweep:

change param Δ luminance Δ red gain subject 0 ΔITA subject 1 ΔITA
exposure −0.50 −14.90 0.000 −36.5 −47.5
exposure −0.25 −8.24 0.000 −20.6 −16.6
exposure −0.10 −3.53 0.000 −8.4 −5.5
exposure +0.10 +3.14 +0.001 +7.2 n/a
white balance 0.97 −0.39 −0.031 −2.2 +4.6
white balance 1.06 +0.78 +0.057 +4.3 −9.2

Slope: about 2.4° of ITA per unit of median luminance, and 90–175° per unit of grey-world red gain (the steeper subject sets the bound). Holding lighting-induced drift near 5° means matching luminance within 2 and red gain within 0.03 — and inside those bounds up to ~5° of any observed drift is still the light, so a complexion finding has to clear it.

Face shape survives lighting#

Variant width/height jaw/width interocular/width
reference 0.78023 0.78055 0.66906
warm ≈3200 K 0.78736 0.77909 0.66837
cool ≈7500 K 0.77791 0.77961 0.67074
exposure +1 stop 0.78352 0.77957 0.66193
exposure −1 stop 0.79287 0.78407 0.66374

Width-over-height spans 1.92%, jaw width 0.6%, interocular 1.3%. Geometry survives light; colour does not, and that is why the two are gated differently. The shipped constant carries 1.93 — rounded up, because understating a noise floor is the direction that makes a gate fire on the light.

Two more measured facts#

A measurement can disappear. Subject 1's complexion was unavailable at +0.1, +0.25 and +1.0 stops while present at +0.5. So "not measured in this frame" is its own outcome and is never folded into "no change" — and the fact that brightening costs coverage unevenly between subjects belongs to 182.C.19.09.

Small faces yield nothing, correctly. Of six faces detected, only the two foreground subjects (relative height 0.163 and 0.158) produced geometry or a complexion. The four bystanders (0.033–0.068) produced neither: the landmarker declined and the skin segment fell under the pixel floor. That matches the 182.C.17 measurability floor.

Two attributes are not measured at all#

Attribute Why
age presentation No age model is staged, and estimating age from face geometry has a documented record of skewing by ethnicity. A number here would be worse than none.
distinctive features Which scar, mole or tattoo matters is a per-subject declaration. Discovering them is a different and more invasive thing than checking a declared one.

Both go to a person, always, and assertHumanReviewComplete refuses sign-off while either is outstanding.

Reproducing#

text
python tools/human-video/evaluator/appearance_attributes.py \
  --image docs/domains/isis/human-video/evidence/2026-09-10-openrouter-face-detection/two-foreground-subjects-plus-bystanders.png

Note: analyse one image per process. Stacking MediaPipe task instances in a single interpreter segfaults reproducibly after a handful, which is why the lighting probe forks per image.

Per-item status this evidence supports#

Item Was Is
182.C.19.02 not-started done; five of seven attributes measured, two routed to human review, colour gated on matched lighting
182.C.19.08 pending this is a second reason thresholds cannot be universal: the same cutoff means different things per lighting
182.C.19.09 pending brightening cost coverage unevenly between the two subjects — a fairness slice with a measurement behind it