MEBeauty is released for NON-COMMERCIAL ACADEMIC RESEARCH ONLY, restricted to facial attractiveness assessment (facial beauty prediction). It must NOT be used for facial identity recognition, verification, surveillance, or re-identification.
These are images of real people. The ratings are subjective human opinions, not measurements of any property of the individuals shown.
By requesting access you agree to all terms below.

MEBeauty: Multi-Ethnic Facial Beauty & Attractiveness Dataset

MEBeauty contains about 2,500 face images rated for attractiveness on a 1-10 scale by people from different social and cultural backgrounds. The dataset is diverse in age and includes both genders and six ethnic groups: Caucasian, Asian, Hispanic, Black, Indian, and Middle Eastern. It provides the average score for each image, all individual attractiveness ratings, and a separate set of personal-preference ratings. This allows researchers to study both generic and personalized facial beauty prediction.

These scores show people's opinions, not facts about the people in the images. They reflect what a particular group of raters found attractive, including their personal tastes, culture, and possible biases. The scores do not measure anyone's value and should not be used to make conclusions about any gender, ethnicity, or age group. A model trained on this dataset learns only the preferences of these raters. The dataset is published to support research on how people judge facial attractiveness.

About this release

A mirror and improved version of the original MEBeauty dataset, published by its first author. The original appeared five years ago; this release keeps the same images and the same ratings — nothing new was collected — and improves everything around them. The ratings, the raters and the image metadata were re-checked with tooling that did not exist in 2021, duplicates and unusable files were removed, and the splits were rebuilt. Counts differ slightly from the original for that reason.

Original repository: https://github.com/fbplab/MEBeauty-database

Terms of use

Non-commercial academic research only.
Facial attractiveness assessment only. Not for face recognition, identification, verification, biometric matching, re-identification or surveillance.
No redistribution. Do not mirror or re-upload the images or ratings.
Citation required (see below).

The images are not owned by the maintainer and remain subject to the rights of their photographers and the people shown. Collection is described in the paper. See LICENSE.

Quick start

Access is gated. Log in and accept the terms once, then:

from datasets import load_dataset

ds = load_dataset("dr-irina-lebedeva/MEBeauty", split="train")     # default config: `fbp`
ds[0]["image"], ds[0]["beauty_score"]          # 256x256 aligned face, 1-10

huggingface-cli login      # first time only

Raters and ratings

Task	Raters	Ratings
Attractiveness	593	68,868
Date preference	631	72,531
Unique raters	860

Config	For
`fbp` (default)	cropped aligned face + score. Minimal setup for training a generic predictor
`fbp_extended`	adds the original image, gender and ethnicity
`personalized_fbp`	adds every individual rating with rater demographics
`personalized_date`	the same for a second task (would the rater date this person)
`full`	everything

The data

Images. Two views of each face: an aligned 256x256 crop (eyes and mouth in the same place every time) and the original image as collected, 400x400 to 5304x6630, with background.

Column	From	Meaning
`image`	`fbp`+	the face image
`image_id`	`fbp`+	identifier for this image, `mebeauty_000001` style
`beauty_score`	`fbp`+	the label. The average rating, with rater leniency removed first
`rating_distribution`	`fbp`+	counts per point on the 1-10 scale, from the raw ratings
`landmarks`	`fbp`+	68 facial points, 136 values, in this config's coordinate space
`split`	`fbp`+	`train`, `val` or `test` under the held-out protocol
`cv_fold`	`fbp`+	cross-validation fold 0-4, covering every image
`image_native`	`fbp_extended`+	the image as collected, at its original resolution, before cropping
`landmarks_native`	`fbp_extended`+	the same 68 points in native coordinates
`gender`	`fbp_extended`+	gender of the person in the image
`ethnicity`	`fbp_extended`+	ethnicity of the person in the image
`plain_mean_score`	`full`+	the simple average of the same ratings, with no correction. For reference — not the score to train on
`score_std`	`full`+	spread of those ratings
`rating_count`	`full`+	how many raters scored this image
`source_url`	`full`+	the source page this image was matched to
`attractiveness_ratings`	`full`+	all individual attractiveness ratings for this image
`date_ratings`	`full`+	all individual personal-preference ratings for this image

Label. beauty_score, on a 1-10 scale, in every config.

Every face was seen by a different group of raters, and raters differ in how generously they score — so a simple average would partly depend on who happened to rate a face. beauty_score removes that effect before averaging. No rater is removed.

The simple average ships too, as plain_mean_score, but only in the full config — for reference, not for training.

Also rating_distribution, a histogram of the raw ratings, for label-distribution learning. 68 facial landmarks ship with every image.

Splits. Two protocols; use one and say which.


80/10/10 fixed split (default)	train 1,962 / validation 250 / test 250
5-fold cross-validation	`cv_fold` 0-4, every image evaluated once

Similar-looking images that may show the same person are grouped so they never cross a split or fold.

from datasets import concatenate_datasets

# cv_fold covers every image, so join the three splits first
data = concatenate_datasets([load_dataset("dr-irina-lebedeva/MEBeauty", split=s)
                             for s in ("train", "validation", "test")])
train = data.filter(lambda r: r["cv_fold"] != 0)
test  = data.filter(lambda r: r["cv_fold"] == 0)

Individual ratings

personalized_fbp, personalized_date and full carry all individual ratings for each image:

Field	Meaning
`rater_id`	identifies a rater within this dataset only
`score`	what this person gave, 1-10
`rater_gender`, `rater_ethnicity`, `rater_age_band`	where known; empty for most raters

More examples

# a different config
ext = load_dataset("dr-irina-lebedeva/MEBeauty", "fbp_extended", split="train")
ext[0]["image_native"], ext[0]["gender"], ext[0]["ethnicity"]

# every individual rating for one face
per = load_dataset("dr-irina-lebedeva/MEBeauty", "personalized_fbp", split="test")
[r["score"] for r in per[0]["ratings"]]        # e.g. [4.0, 7.0, 6.0, ...]
per[0]["beauty_score"]                         # the label for that face

# train on one subgroup
asian_women = ext.filter(lambda r: r["ethnicity"] == "asian"
                                and r["gender"] == "female")

Version history

2.1.0 — new label, beauty_score, which corrects for raters who score high or low in general. The old score_mean has the same values but a new name, plain_mean_score, and now sits in the full config only. Results from 2.0.0 are still valid — just say which score you used, since the two can differ by up to 1.2 for one face.

2.0.0 — first Hugging Face release.

Limitations

Attractiveness ratings are subjective opinions. A model trained here predicts what these raters said, not a property of anyone shown.
Labels were recomputed from the individual ratings, so numbers from papers using the 2022 release are not directly comparable.

Citation

@article{lebedeva2022mebeauty,
  title   = {MEBeauty: a multi-ethnic facial beauty dataset in-the-wild},
  author  = {Lebedeva, Irina and Guo, Yi and Ying, Fangli},
  journal = {Neural Computing and Applications},
  volume  = {34},
  number  = {17},
  pages   = {14169--14183},
  year    = {2022},
  doi     = {10.1007/s00521-021-06535-0}
}

Contact

Irina Lebedeva, PhD dr.irina.lebedeva@gmail.com https://www.irina-lebedeva.com

Downloads last month: 61

Total file size:

762 MB