30,000 celebrity faces dissected into 19 labeled layers

What it does CelebAMask-HQ is a dataset repository—not a training framework—hosting 30,000 high-resolution portraits drawn from the CelebA-HQ lineage. Each 512×512 image carries a manually annotated segmentation mask split into 19 classes, covering everything from skin and eyebrows to accessories like hats, earrings, and necklaces. Researchers use it to train and benchmark face-parsing networks, recognition systems, and generative models that manipulate facial features.

The interesting bit The annotations are almost surgically granular: the masks distinguish hair from hats and lips from mouth interiors, which is precisely the control signal needed for interactive facial editing. The README notes that identity and attribute labels are not included here; you must request those separately from the original CelebA team, so the repo is strictly a delivery vehicle for images and masks.

Key highlights

• 30,000 images at 512×512, manually annotated rather than auto-generated
• 19 segmentation classes spanning facial anatomy and accessories (skin, nose, eyes, hair, hat, eyeglass, earring, necklace, neck, cloth, etc.)
• Derived from CelebA-HQ, inheriting the same high-quality upscaled face portraits
• Explicitly intended for face parsing, recognition, and GAN-based generation or editing
• Redistribution is limited to internal single-site use; broader publication or distribution is prohibited

Caveats

• Strictly non-commercial license: any commercial use, resale, or exploitation of the images or derived data is forbidden
• Identity and attribute labels require a separate request to the CelebA team, so the repo does not ship a complete metadata package

Verdict Worth downloading if you are building face-parsing or GAN-editing research inside academic or non-commercial walls. Skip it if you need commercial licensing or a turnkey dataset with full identity and attribute labels out of the box.