README.md

metadata

license: cc-by-nc-4.0
pretty_name: HairCS
size_categories:
  - 10K<n<100K
tags:
  - hair
  - strands
  - 3d
  - computer-graphics
  - point-cloud

HairCS

HairCS is a strand-based hair dataset: 53,336 hair assets in 13 versions, each stored as a point cloud of 60,000 strands x 64 samples (xyz, float16), plus a rendered preview image of every asset, per-asset hairstyle labels, the head/scalp geometry the hair is grown on, and a small Python toolkit (viewer and strand-to-surface conversion).

All versions share the same head and the same coordinate frame (see meta/).

Versions

Version	Assets	Description
v0	1,076	Base hair
v1	1,076	v0 + scale and fuzz modifier
v2	818	Curly profile 1
v3	844	Curly profile 2
v4	650	Curly + scale and fuzz modifier
v5	1,539	Blend: bob (00000-00065) and shoulder (00066-01538)
v6	1,539	v5 + curly profile 1
v7	6,086	Blend: shoulder
v8	6,086	v7 + curly profile 1
v9	8,585	Blend: short
v10	8,585	v9 + curly profile 1
v11	8,226	Blend: long
v12	8,226	v11 + curly profile 1

Terminology:

• Blend – assets produced with the hair blending method described in the paper.
• Curly profile 1 – the helix operator from the paper, with its first parameter profile.
• Curly profile 2 – the helix operator with a second parameter profile.
• Scale and fuzz modifier – per-strand scaling plus a fuzz perturbation applied on top of the base hair.

Repository layout

data/vN/XXXXX.npz        one asset per file, key "P": (60000, 64, 3) float16, samples ordered root -> tip
render/vN.zip            one PNG preview per asset, named XXXX.png (4-digit index, same order as data/vN)
label/vN.csv             hairstyle class per asset (columns: number,class)
meta/                    head and scalp geometry shared by all versions (OBJ)
  full_body_geo.obj
  head_water_tight_boundary.obj
  head_with_uv.obj
  scalp_with_uv.obj
argument/                viewer + strand-to-surface tools (see argument/README.md)

Every data/vN/ folder holds exactly the number of assets listed in the table, numbered 00000 to N-1.

Loading an asset

import numpy as np

d = np.load("data/v0/00000.npz")
P = d["P"].astype(np.float32)     # (60000, 64, 3): 60000 strands, 64 samples per strand, xyz
roots, tips = P[:, 0], P[:, -1]   # samples are ordered root -> tip

The head meshes in meta/ are in the same coordinate frame as the strands, so an asset can be placed on head_with_uv.obj or scalp_with_uv.obj without any transform.

Labels

label/vN.csv gives one hairstyle class per asset:

Class	Meaning
short	short hair
bob	bob cut
shoulder	shoulder-length hair
long	long hair
gather	gathered / tied-back hair

For v0-v6 the classes were assigned per asset. For v7-v12 every asset of a version shares the class of the blend target of that version (v7/v8: shoulder, v9/v10: short, v11/v12: long).

Downloading

The full dataset is about 963 GB. Fetch only what you need with --include:

pip install -U huggingface_hub
hf download HairCS2027/HairCS --repo-type dataset --local-dir HairCS \
    --include "meta/*" "label/*" "argument/*"            # small files, ~25 MB
hf download HairCS2027/HairCS --repo-type dataset --local-dir HairCS \
    --include "data/v4/*" "render/v4.zip"                 # one version, ~11 GB

Folder	Size
data/	~932 GB (13 versions, 10 to 154 GB each)
render/	~31 GB (13 zips, 0.3 to 7 GB each)
meta/, label/, argument/	< 25 MB

Tools

argument/ contains a strand viewer (run_viewer.py) and a tool that wraps an asset in a watertight mesh (run_strand2surface.py, with a batch runner and optional GPU path). Usage and the expected on-disk layout are described in argument/README.md.

License

CC BY-NC 4.0. Non-commercial use only.

Citation

The accompanying paper is under review. A citation entry will be added on publication.