The darkest face in the render comes back wrong. A new study on arXiv puts a number on how wrong: 49.43 degrees of median skin tone error for the darkest ITA category, against 12.12 for the lightest. That is a fourfold gap in colour fidelity, measured across 19,848 MetaHuman renders in Unreal Engine. Every brand deploying virtual try-on or AI-generated campaign imagery inherits the distortion unless it calibrates.

The study tested four extraction methods for pulling skin colour from photographs before mapping it onto MetaHuman avatars. The best performer combined illumination compensation (TRUST) with multidimensional colour masking, achieving a median ΔE of 3.77 across all tones. For the darkest skin category, the best pipeline still yielded an ITA error of 37.34 degrees. Without illumination correction, error climbed past 56 degrees and the pipeline compressed darker phenotypes toward intermediate ITA classes rather than preserving their original category.

Lighting compounded the problem. Under standard frontal illumination, the darkest tones hit 53.85 degrees of error. Reconstructing lighting matched to the original photograph dropped that to 11.37 — most of the damage is fixable, if anyone bothers to match.

The pattern is Kodak’s Shirley card for the rendering era.

From the 1950s through the late 1970s, Kodak calibrated its film chemistry to a single white reference model. Darker skin rendered flat, muddy, stripped of detail. The company recalibrated only when furniture and chocolate manufacturers complained their products looked wrong on film — not when people did. The virtual try-on industry faces the same structural incentive: pipelines tuned to a default range, with failure modes invisible to the teams shipping them.

A separate study presented at WACV 2025 found the same bias pattern in 3D relightable face generators, tracing the root cause to lighting estimation methods that systematically underestimate illumination on dark skin. The spherical harmonics coefficients extracted from dark-skinned faces formed a distinct cluster, skewed toward dimmer lighting values. The distortion is not confined to one pipeline or one rendering engine. It is an inherited defect in how these systems extract and represent colour from photographs.

What the study gives the industry is a benchmark, not a product. A brand deploying MetaHuman-based try-on can now measure its pipeline against published ΔE and ITA thresholds per skin tone category. The fixes are known: matched lighting cut the darkest-tone error from 53.85 to 11.37 degrees; illumination-compensated extraction reduced the uncorrected baseline from approximately 56 degrees to 37 — a cut of roughly a third. The question is whether brands will run the calibration before a customer screenshots the proof that they didn’t.