COMPUTATIONAL PHOTOGRAPHY · PORTRAIT PERCEPTION
FaceContrast
1Université Paris-Saclay, CNRS, ENS Paris-Saclay, Centre Borelli, France
2DXOMARK, France 3Lingnan University, Hong Kong SAR, China
THE IDEA
Overall tonal separation, eye definition, and fine skin detail contribute differently to how a portrait looks. We measure each component and validate it against human perception.
The quality of smartphone portrait photography has become increasingly important to both manufacturers and users. Recent devices use complex computational pipelines to improve facial rendering under varied lighting. These pipelines vary widely across manufacturers, so the same portrait can look noticeably different across smartphones. These differences affect multiple aspects of facial appearance, making perceptual preference difficult to characterize with a single measure. To address this, we propose an interpretable framework that decomposes face contrast into three components: overall facial tonal separation, eye‑region contrast, and micro‑contrast of fine skin details. We formulate dedicated metrics for each component and validate them via pairwise human judgments, demonstrating that our approach correlates more closely with human contrast perception than traditional contrast measures. To support this analysis, we introduce a realistic mannequin‑based dataset designed for systematic facial contrast evaluation, comprising 508 full‑resolution captures from 29 smartphones under diverse lighting, with region‑specific masks and extensive pairwise perceptual annotations. As a downstream analysis, we show that combining the three component‑wise metrics provides a meaningful basis for modeling balanced facial contrast preference in a controlled setting, and we present a preliminary evaluation on a smaller real‑portrait dataset indicating that insights derived from the mannequin study can extend to real images. Our work provides a controlled first step toward defining and evaluating face contrast in portrait imaging.
01 / METHOD
We define three components of contrast and study how their balance relates to portrait preference.
CONTRAST AMOUNT · FACE
The global tonal separation across the face: how clearly bright and dark regions are separated. It describes facial shading and the transitions that give the face depth.
How strongly are light and shadow separated across the face?
CONTRAST AMOUNT · EYES
The local lightness separation within the eye region, between dark structures such as the pupil and brighter structures such as the sclera. It describes how distinctly these structures appear.
How clearly do dark and bright structures stand apart within the eyes?
CONTRAST AMOUNT · SKIN DETAIL
The visibility of fine facial details, such as skin pores and small local structures. It describes how strongly fine skin texture is rendered in the final portrait.
How visible and pronounced are the fine details of the skin?
PERCEPTUAL PREFERENCE · THE THREE COMPONENTS TOGETHER
The perceived harmony among overall tonal separation, eye definition, and fine-detail visibility: whether they form a natural and visually pleasing balance.
Which portrait has the more pleasing balance of facial contrast?
More contrast is not always better. Balanced contrast is a separate preference judgment over the three components, rather than a fourth measure of contrast amount.
All three measures use the L* (lightness) channel of CIELAB, normalized to [0, 1] and denoted by L. This separates tonal variation from chromatic variation. Select a component to explore its computation.
Overall contrast captures tonal transitions across the face. Local lightness variation is normalized by the local mean and pooled over five spatial scales, from fine shading to broader facial structure.
Here, μr and σr are the local mean and standard deviation of L, the normalized CIELAB L* channel. The response is averaged over the face and across window sizes 16, 32, 64, 128, and 256.
COMPUTATION AT A GLANCE
Five window sizes, one face-wide measure.
Eye contrast measures the separation between dark and bright structures inside the eyes. A single Otsu threshold partitions lightness values from both eye masks into two classes.
Here, μbright and μdark are class means of L, the normalized CIELAB L* channel, within the merged left and right eye masks.
COMPUTATION AT A GLANCE
Separate lightness. Compare the means.
Lightness values from both eye regions
Eye contrast = μbright − μdarkA larger difference indicates stronger eye contrast.
Micro-contrast measures fine skin texture while reducing the influence of broad shading. We extract texture in log-lightness space, measure its local energy, and compress strong responses before pooling over the skin.
Here, Ωs is the skin mask. The thresholds q10 and q90 are the 10th and 90th percentiles of Slocal within that mask. Averaging the retained responses gives Cmicro.
COMPUTATION AT A GLANCE
Isolate texture. Measure its visible strength.
02 / DATASET
Two realistic mannequins, Eugene and Diana, let us compare rendering while keeping facial structure and surface texture fixed.
Original captures and contrast-adjusted variants form an 800-image annotation set. Trained image-quality experts completed four separate comparison tasks over 160 hours.
Pairwise judgments are converted to Just Objectionable Difference (JOD) scores. A difference of 1 JOD corresponds to a 75% choice probability for the higher-scoring image.
Dataset access
A download link has not yet been added to this project page. See the project repository ↗ for available resources.
03 / RESULTS
Correlation with component-specific human judgments, without fitting the three contrast metrics to the annotations.
| Proposed measure | Within-lighting | Mixed-lighting | ||
|---|---|---|---|---|
| PLCC ↑ | SRCC ↑ | PLCC ↑ | SRCC ↑ | |
| Overall contrast | 0.792 | 0.780 | 0.688 | 0.693 |
| Eye contrast | 0.879 | 0.855 | 0.892 | 0.891 |
| Micro-contrast | 0.849 | 0.842 | 0.797 | 0.787 |
Within-lighting: device-disjoint evaluation on Groups 1–4. Mixed-lighting: evaluation on Group 5. Values reproduced from the main paper; see the paper for baselines and evaluation details.
BEYOND THE LAB
A preliminary evaluation uses 444 images across 10 subjects/scenes, with varied facial appearances, poses, and device processing.
The proposed zero-shot metrics achieve SRCC of 0.825 for overall contrast, 0.768 for eye contrast, and 0.731 for micro-contrast, macro-averaged across scenes.
Broader evaluation across identities, skin tones, poses, and capture conditions remains future work.
REFERENCE
Accepted at WACV 2027. Available on arXiv:2610.26517.
@misc{cho2026facecontrast,
title = {Interpretable Face Contrast Modeling for
Smartphone Portrait Rendering},
author = {Cho, Yujin and Pochon, Beno{\^i}t and
Morel, Jean-Michel and Facciolo, Gabriele},
year = {2026},
eprint = {2610.26517},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.26517}
}This work was financed by ANRT grant N° 2023/0335 and used GENCI-IDRIS HPC and storage resources on Jean Zay under allocations AD011017622, AD011014895R2, AD011014895R1, and AD011014895R.