Layered Labs

DatasetsSeptember 2026

Grade acne severity and find lesions against one checked source.

0
facial photographs
0
lesion boxes
0
original cross-validation folds

1Summary

ACNE04 is 1,457 facial photographs graded for acne severity, with 18,983 lesion boxes, rebuilt from the original ACNE04 release into one dataset. Every image carries its severity grade, its lesion boxes and its place in each of the original five cross-validation folds, so results can be compared directly with the published benchmark.

2Motivation

Each tile is ten photographs. Some earlier copies held barely a quarter of the dataset.

ACNE04 was first released as two separate archives: one with the images and their severity grades, the other with the lesion boxes. Using both meant downloading each, matching files by name, and joining them by hand.

Copies on Hugging Face did not solve this. Some held as few as 394 of the 1,457 images, some dropped the lesion boxes, and some reused the labels for unrelated tasks with no severity grade left. A model tested on one of these copies cannot be compared with the original benchmark.

3Composition

The four Hayashi severity grades, with the number of photographs in each. Faces are drawn for illustration.

One row per photograph: the image, its severity grade on the four-level Hayashi scale, the original lesion-count label, the number of lesions boxed, the boxes themselves, and its train or test membership in each fold.

SeverityLevelImagesShare
Mild051335%
Moderate163343%
Severe218212%
Very severe31299%
Images per Hayashi severity grade, across all 1,457 photographs.

Moderate acne is the most common grade and very severe the least, so a model that always guesses moderate is right 43% of the time.

4Collection

The two original archives, joined row by row.

The images and labels come unchanged from the official release linked from the authors' repository (Wu et al., ICCV 2019). The images and severity grades come from its classification archive, and the lesion boxes from its detection archive, in Pascal VOC format. The two were joined row by row on the file name.

5Cleaning and labels

The columns of each row. The severity level is the primary label.

Nothing was re-annotated or re-graded. The severity grades and lesion boxes are the original authors' annotations.

  • Severity is the primary label: level from 0 to 3, and severity as a word (mild, moderate, severe, very severe).
  • Lesion-count label is kept as released. It mixes lesion count with severity in a way the original authors did not fully document, so use level for severity.
  • Folds. The five test folds are close to, but not exactly, a partition of the images: a few images appear in more than one test fold. This comes from the original release. The canonical split column uses fold 0 only, which is a clean split.

6Uses

The canonical fold 0 split holds out 292 of the 1,457 images for testing.

For training and testing acne severity classifiers and lesion detectors, and for testing general multimodal models on dermatology grading against the Hayashi ground truth.

To compare with the published benchmark, report results on all five folds. For a single quick split, use the canonical split column: 1,165 training images and 292 test images.

It is a research dataset. It is not for diagnosis or for decisions about real patients.

7Distribution

One call loads every image with its labels.

On Hugging Face as Layered-Labs/ACNE04, with the images stored inside the files, so one call loads everything:

from datasets import load_dataset
ds = load_dataset("Layered-Labs/ACNE04")

The original authors released ACNE04 for academic research only, and this release keeps that restriction. It is not for commercial use. For other uses, contact the original authors through their repository.

8Maintenance

Every lesion carries the box the original annotators drew; 18,983 boxes across the dataset. Drawn for illustration.

Layered Labs maintains this release. Questions and corrections go to the dataset's Community tab on Hugging Face.

Living draft, updated November 2026. This page can change before formal publication.