MEBeauty: Multi-Ethnic Facial Beauty Benchmark

• MEBeauty is a multi-ethnic facial beauty dataset that provides dense, continuous attractiveness scores across six race groups for robust bias analysis.
• It serves as a benchmark for evaluating deep beauty regression models, makeup recommendation systems, and other aesthetic assessment pipelines.
• Recent studies leveraging MEBeauty reveal significant cross-dataset performance variations and highlight challenges in achieving fairness across demographics.

Searching arXiv for papers on MEBeauty, facial beauty prediction, and related makeup recommendation/transfer. MEBeauty denotes two closely related objects in the recent literature. In the narrow sense, it is a labeled, multi-ethnic facial beauty dataset introduced in the form cited by Lebedeva et al. and used as a core benchmark for deep beauty regression and bias analysis. In a broader sense, later work uses “MEBeauty” or “MEBeauty-style” to describe machine evaluation or estimation of beauty, including continuous facial attractiveness prediction, personalized makeup recommendation, virtual try-on, and beauty-oriented product retrieval (Hamel et al., 29 Sep 2025, Boukhari, 21 Sep 2025).

1. Definition and terminological scope

In the dataset sense, MEBeauty is defined as follows: “The MEBeauty dataset comprises 2,550 images representing Black, Asian, Caucasian, Hispanic, Indian, and Mideastern female and male faces. Each image is rated on a [1, 10] beauty scale by approximately 300 individuals from diverse cultural and social backgrounds.” Its labels therefore encode a continuous beauty target with unusually dense human annotation relative to earlier facial beauty benchmarks (Hamel et al., 29 Sep 2025).

The acronym itself is not explicitly expanded in the bias-analysis paper that uses it, and that paper states that it is not framed as “Middle Eastern only”; rather, it is treated as a multi-ethnic dataset spanning six race categories. This distinction matters because MEBeauty has sometimes been misunderstood as a geographically restricted corpus, whereas in the cited usage it is explicitly broader and is operationalized as a race-labeled benchmark for multi-group analysis (Hamel et al., 29 Sep 2025).

A second, broader usage appears in later facial beauty prediction and makeup-system papers. There, “MEBeauty” functions as a field label for machine beauty assessment or beauty-oriented face technologies. Under that broader reading, continuous facial beauty regression, aesthetic representation learning, personalized makeup synthesis, and product-attribute extraction all fall within the same technical space (Boukhari, 21 Sep 2025).

2. Dataset design, labels, and comparative position

MEBeauty occupies an intermediate position between classical portrait-based beauty benchmarks and later application-specific datasets. It is smaller than SCUT-FBP5500 and much smaller than FairFace or LiveBeauty, but it offers broader ethnic coverage than SCUT-FBP5500 and denser annotation per image than either SCUT-FBP5500 or LiveBeauty (Hamel et al., 29 Sep 2025, Li et al., 5 Jan 2025).

Dataset	Images and labels	Distinctive properties
MEBeauty	2,550 images; beauty scores in $[1,10]$; $\sim 300$ raters	Six ethnic categories; female and male faces
SCUT-FBP5500	5,500 images; beauty scores in $[1,5]$; 60 raters	Larger; four demographic groups concentrated in Asian and Caucasian subjects
LiveBeauty	10,000 images; MOS in $[1,5]$; 20 raters	Live-streaming domain; Asian-only; 200,000 total annotations

Relative to SCUT-FBP5500, MEBeauty has fewer images, more ethnic diversity, a broader rating scale, and many more ratings per image. SCUT-FBP5500 contains 2,000 Asian females, 2,000 Asian males, 750 Caucasian females, and 750 Caucasian males, with beauty scores averaged from 60 raters on a $[1,5]$ scale. MEBeauty instead spans Black, Asian, Caucasian, Hispanic, Indian, and Middle Eastern categories, covers both female and male faces, and uses a $[1,10]$ scale with approximately 300 raters per image (Hamel et al., 29 Sep 2025).

FairFace, by contrast, is not a beauty-labeled dataset in this context. It is used only as an unlabeled, demographically balanced test set with 86,744 images and seven racial groups. This makes MEBeauty especially valuable when the task requires both demographic labels and beauty ground truth, rather than demographic balance alone (Hamel et al., 29 Sep 2025).

Later benchmark construction underscores MEBeauty’s status as a reference dataset rather than a domain-complete solution. LiveBeauty, for example, explicitly contrasts its 10,000-image live-streaming corpus with MEBeauty’s 2,550-image static multi-ethnic setup and evaluates newer models on both, treating MEBeauty as one of the standard public FAP benchmarks (Li et al., 5 Jan 2025).

3. Preprocessing and use in regression pipelines

In the bias-analysis study, MEBeauty serves as both a training source and an evaluation target for deep facial beauty regression. Preprocessing is shared with other datasets: metadata tables are constructed with filename, race label, and beauty score; scores are normalized to $[0,1]$; faces are cropped with MTCNN and padded to square images; and pixels are normalized with ImageNet statistics, using $\mu=[0.485,0.456,0.406]$ and $\sigma=[0.229,0.224,0.225]$ (Hamel et al., 29 Sep 2025).

For MEBeauty scores in $[1,10]$, the paper applies linear normalization of the form

$\sim 300$0

with $\sim 300$1 and $\sim 300$2. No additional quality-control protocol beyond face detection and normalization is specified (Hamel et al., 29 Sep 2025).

The model trained on MEBeauty is a fine-tuned ResNet-152 pretrained on ImageNet, with the final fully connected layer replaced by a single-output linear layer predicting a scalar normalized beauty score in $\sim 300$3. Training uses MSE, Adam with $\sim 300$4, ReduceLROnPlateau with factor $\sim 300$5 and patience of 2 epochs, and early stopping with patience 5 on validation MSE. The schedule proceeds in three phases: training only the final layer, then unfreezing Conv5, then unfreezing Conv4. Data are split into train $\sim 300$6, validation $\sim 300$7, and test $\sim 300$8, with batch size 32 for MEBeauty (Hamel et al., 29 Sep 2025).

Under this protocol, the MEBeauty-trained model achieves test MSE $\sim 300$9 on held-out MEBeauty data and cross-dataset MSE $[1,5]$0 when evaluated on SCUT-FBP5500. The corresponding SCUT-trained model achieves test MSE $[1,5]$1 on SCUT and cross-dataset MSE $[1,5]$2 on MEBeauty. This suggests that dataset-specific label scales, demographics, and aesthetics materially affect cross-dataset transfer (Hamel et al., 29 Sep 2025).

MEBeauty is also used as a benchmark for newer architectures. In LiveBeauty, the proposed FPEM model reports, on MEBeauty with an 80/20 split, SROCC $[1,5]$3, PLCC $[1,5]$4, and KROCC $[1,5]$5, outperforming the compared baselines in that study (Li et al., 5 Jan 2025).

4. Fairness analysis and the refutation of “diversity implies fairness”

The most consequential recent discussion of MEBeauty concerns bias. In the cited study, MEBeauty is one of the two core datasets used to examine ethnicity-based bias in deep learning facial beauty prediction, and it functions as both a training source and a labeled multi-group testbed (Hamel et al., 29 Sep 2025).

When a SCUT-trained model is applied to MEBeauty, substantial inter-group disparities appear in both predicted scores and signed errors. Mean predicted scores are $[1,5]$6 for Asian faces, $[1,5]$7 for Black faces, $[1,5]$8 for Caucasian faces, $[1,5]$9 for Hispanic faces, $[1,5]$0 for Indian faces, and $[1,5]$1 for Middle Eastern faces. Mean signed errors are $[1,5]$2 for Asian faces, $[1,5]$3 for Black faces, $[1,5]$4 for Caucasian faces, $[1,5]$5 for Hispanic faces, $[1,5]$6 for Indian faces, and $[1,5]$7 for Middle Eastern faces. Middle Eastern faces thus show the largest positive mean error, while Asian and Indian categories are underestimated on average (Hamel et al., 29 Sep 2025).

These disparities are statistically validated with non-parametric tests. For $[1,5]$8 groups, the Kruskal–Wallis statistic is

$[1,5]$9

On MEBeauty, the reported statistics are $[1,5]$0 for prediction distributions and $[1,5]$1 for error distributions, with $[1,5]$2 in both cases. Post hoc Dunn tests with Benjamini–Hochberg correction further show that only 2 of 15 pairwise race comparisons, or 13.3%, satisfy distributional parity, and only 3 of 15, or 20%, satisfy error parity (Hamel et al., 29 Sep 2025).

Cross-dataset results sharpen the point. When the MEBeauty-trained model is tested on SCUT-FBP5500, mean predictions are similar across Asian and Caucasian groups, but Asian mean error is $[1,5]$3 and Caucasian mean error is $[1,5]$4. Mann–Whitney U and Kolmogorov–Smirnov tests reject both prediction parity and error parity. On FairFace, Kruskal–Wallis tests yield $[1,5]$5 for the SCUT-trained model and $[1,5]$6 for the MEBeauty-trained model, both with $[1,5]$7; only 1 of 21 race pairs, or 4.8%, satisfies distributional parity for the SCUT model, and only 2 of 21, or 9.5%, for the MEBeauty model (Hamel et al., 29 Sep 2025).

A common misconception is therefore directly contradicted by the reported evidence: broader ethnic representation in the training data does not automatically yield a fairer beauty predictor. The authors explicitly state that their findings refute the initial hypothesis that MEBeauty’s broader ethnic representation would yield fairer models than SCUT-FBP5500 (Hamel et al., 29 Sep 2025).

5. MEBeauty as a broader research program

Outside the strict dataset sense, “MEBeauty” has come to denote a wider technical program centered on machine beauty estimation and beauty-oriented face systems. In facial beauty prediction, this includes high-capacity regressors such as MD-Net, which combines a frozen Stable Diffusion v1.5 U-Net encoder with Vision Mamba and cross-attention fusion, achieving PC $[1,5]$8, MAE $[1,5]$9, and RMSE $[1,10]$0 on SCUT-FBP5500 (Boukhari, 21 Sep 2025). A related heterogeneous ensemble, VM-BeautyNet, fuses a Vision Transformer with Vision Mamba and reports PC $[1,10]$1, MAE $[1,10]$2, and RMSE $[1,10]$3 on the same benchmark (Boukhari, 17 Oct 2025).

The broader MEBeauty space also includes recommendation and synthesis systems. “Smart Mirror” learns relations among makeup-invariant facial features, facial attributes, and makeup attributes with a latent SVM based on Wang and Mori, using a 500-image Beauty Makeup Dataset and rendering foundation, eye shadow, and lip color through guided filtering and region-wise blending (Nguyen et al., 2017). DMT, a disentangled makeup transfer GAN, separates identity code $[1,10]$4 from makeup code $[1,10]$5, supports pairwise transfer, interpolated strength control, multiple-reference fusion, and sampling from a Gaussian prior, and reports a user-study preference of 57.3% over BeautyGAN as well as substantially better reconstruction metrics (Zhang et al., 2019).

A further extension links beauty products, retrieval, and try-on. The material-property extraction pipeline for enhanced beauty product discovery uses six specialized models for best-image selection, format classification, shade detection, base-color regression, finish classification, and reflective-color regression. On eyeshadow, it reports selection accuracy 92.22%, format F1 around 0.90, shade-detection mAP around 0.93, finish F1 around 0.84–0.89, and base-color $[1,10]$6 around 4.86–6.35, then uses these extracted attributes for recommendation and virtual try-on (Dezaki et al., 2023).

This broader usage suggests that MEBeauty is no longer only a dataset label. It is also an umbrella for a research trajectory in which beauty is operationalized as a continuous human-aligned signal, a structured cosmetic recommendation task, or a multimodal retrieval-and-rendering problem.

6. Limitations, ethical concerns, and likely directions

MEBeauty’s central methodological contribution is not that it solves fairness, but that it makes fairness diagnosable. Its six-group structure enables 15 pairwise comparisons under non-parametric parity tests, and its race labels permit analyses that are impossible on unlabeled beauty datasets (Hamel et al., 29 Sep 2025).

At the same time, the evidence indicates clear limitations. The bias-analysis paper attributes persistent disparities to sampling bias and labeling bias, noting that MEBeauty’s label distributions across ethnic groups are skewed and that even $[1,10]$7 raters per image do not guarantee demographically balanced or culturally neutral judgments. This suggests that annotator diversity alone does not neutralize systemic preferences encoded in beauty labels (Hamel et al., 29 Sep 2025).

The broader literature reaches similar conclusions. High-performing predictors such as MD-Net explicitly caution that they approximate one dataset’s rating distribution rather than an objective or universal notion of beauty, while LiveBeauty notes the need for fairness and cross-cultural expansion, especially because its own live-streaming benchmark is Asian-only (Boukhari, 21 Sep 2025, Li et al., 5 Jan 2025). A plausible implication is that future MEBeauty work will need jointly optimized dataset curation, multi-dataset validation, and fairness-aware training, rather than relying on demographic breadth as a proxy for equity.

Accordingly, the most concrete recommendations now attached to MEBeauty are procedural rather than celebratory: enforce demographic balance via stratified sampling, implement annotator diversity quotas, validate distributional parity and error parity before deployment, and report sampling biases and labeling processes transparently. In that sense, MEBeauty has become both a resource and a warning: a resource because it enables rigorous multi-group analysis, and a warning because even comparatively diverse beauty datasets can still encode and amplify narrow aesthetic norms (Hamel et al., 29 Sep 2025).