README.md

metadata

license: mit
task_categories:
  - image-classification
  - text-to-image
tags:
  - synthetic
  - faces
  - face-recognition
  - fairness
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files: data/*.tar

SynthFaces: balanced synthetic faces

Synthetic face dataset from the paper SynthFaces (accepted to BMVC 2026), arXiv:2610.40119. Code: ControlGenAI/SynthFaces. Identity sampler weights: shredder67/id-flow.

Contents

• 15,054 synthetic identities, 10 images each: 150,540 images, 512x512 PNG.
• <id>_0 is the base face generated from the sampled identity embedding (ID-Flow, then Arc2Face). <id>_1 to <id>_9 are variations of the same identity.
• Every identity has a label [age, race, gender] (FairFace categories) shared by its 10 images.

Format

WebDataset tar shards in data/ (250 identities per shard). Each sample consists of:

• <id>_<k>.png: the image
• <id>_<k>.json: {"identity", "age", "race", "gender", "age_name", "race_name", "gender_name"}

labels.jsonl in the repo root has the same labels for all 150,540 images (field image is <id>/<id>_<k>.png).

from datasets import load_dataset
ds = load_dataset("shredder67/synthfaces-faces-balanced", split="train", streaming=True)
sample = next(iter(ds))
sample["png"], sample["json"]

Label vocabularies

• age: 0-2, 3-9, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+ (ids 0 to 8)
• race: White, Black, Latino_Hispanic, East Asian, Southeast Asian, Indian, Middle Eastern (ids 0 to 6)
• gender: Male, Female (ids 0, 1)

Label counts (images)

age	count	race	count	gender	count
0-2	2,770	White	25,580	Male	68,870
3-9	22,290	Black	18,500	Female	81,670
10-19	25,690	Latino_Hispanic	24,510
20-29	21,540	East Asian	20,700
30-39	19,690	Southeast Asian	21,260
40-49	19,500	Indian	18,790
50-59	18,090	Middle Eastern	21,200
60-69	15,520
70+	5,450

Citation

See the GitHub repository for the BibTeX entry.