TailorRig

Human character rigging dataset: 39,073 humanoid meshes with skeletons and skinning weights, stored as .npz files and packed into four source-based subset archives.

Suitable for training and evaluation of skeleton generation (auto-rigging), skinning-weight prediction (auto-skinning), joint-position regression, and related tasks.

Links

• Paper: https://arxiv.org/pdf/2611.02902
• Code: https://github.com/Tencent/TailorRig
• Project page: https://xxxx0601.github.io/trig

Dataset Structure

Subset (directory)	File	Archive size	Samples	Notes
rignet	rignet/npz_public.zip	121 MB	715	RigNet-style skeletons, 11–54 joints (median 22), including extra bones for skirts/clothing (e.g. pelvis_skirt_01_01)
humanrig	humanrig/npz_public.zip	14.1 GB	11,419	Standard human skeletons, higher poly count (sampled 71,810 vertices / 27,290 faces / 22 joints)
articulation_xlv2	articulation_xlv2/npz_public.zip	15.9 GB	23,452	High-subdivision skeletons (sampled 35,833 vertices / 57,042 faces / 58 joints); uuid is a UUID string
vroid	vroid/npz_public.zip	4.0 GB	3,487	Created in-house by the authors with VRoid Studio (sampled 16,813 vertices / 26,701 faces / 94 joints); uuid is numeric

• Total: 39,073 samples, about 34 GB (Git LFS / Xet storage)
• Each .npz corresponds to one character; the internal path is npz_public/<uuid>.npz
• All four subsets share an identical schema, so they can be mixed directly for training
• The vroid subset was produced by us using VRoid Studio (not collected from external sources); the other three subsets are derived from their respective named sources

npz Field Reference

Field	shape	dtype	Description
vertices	(V, 3)	float32	Vertex coordinates
faces	(F, 3)	int64	Triangle face indices
vertex_normals	(V, 3)	float32	Per-vertex normals
face_normals	(F, 3)	float32	Per-face normals
joints	(J, 3)	float32	Joint world coordinates (rest pose)
root_index	()	int64	Root joint index (always 0 in this dataset)
joint_names	(J,)	<U	Joint names, e.g. pelvis / spine_01 / thigh_r
parents	(J,)	int64	Parent joint index; the root is -1
matrix_local	(J, 4, 4)	float32	Local matrix of each joint relative to its parent
body_joint_mask	(J,)	bool	Whether the joint is a main-body joint (excludes extra bones such as clothing/skirts)
skinning_weights_value	(nnz,)	float32	Values of the sparse (COO) skinning-weight matrix
skinning_weights_row	(nnz,)	int64	COO row (vertex index)
skinning_weights_col	(nnz,)	int64	COO col (joint index)
skinning_weights_shape	(2,)	int64	Sparse matrix shape (V, J)
uuid	()	<U	Unique sample ID (numeric for rignet/humanrig/vroid, a UUID for articulation_xlv2)

Full statistics of the rignet subset (measured over all 715 samples):

Metric	min	median	max
Vertices V	1,592	3,084	9,781
Faces F	3,057	5,996	20,088
Joints J	11	22	54
Skinning non-zeros nnz	1,933	4,684	14,148
Skinning density nnz/(V·J)	2.2%	6.5%	13.1%

Usage

1. Download and read a single sample

import io
import zipfile
import numpy as np
from huggingface_hub import hf_hub_download
from scipy.sparse import coo_matrix

ZIP = "rignet/npz_public.zip"          # or humanrig/, articulation_xlv2/, vroid/
path = hf_hub_download("tencent/TailorRig", ZIP, repo_type="dataset")

with zipfile.ZipFile(path) as z:
    name = "npz_public/11712.npz"      # use z.namelist() to list all samples first
    d = np.load(io.BytesIO(z.read(name)), allow_pickle=True)

V = d["vertices"]                      # (V, 3)
F = d["faces"]                         # (F, 3)
J = d["joints"]                        # (J, 3)
parents = d["parents"]                 # (J,)

# Reconstruct the skinning-weight matrix (V, J)
W = coo_matrix(
    (d["skinning_weights_value"], (d["skinning_weights_row"], d["skinning_weights_col"])),
    shape=tuple(d["skinning_weights_shape"]),
).tocsr()
print(W.shape, W.nnz, d["joint_names"])

Note: hf_hub_download downloads the entire subset archive (up to 15.9 GB). If you only need a few samples, fetch each subset separately, or use HfFileSystem with HTTP Range requests to pull only the target entries inside the zip.

2. Iterate over all samples in a subset

import io, zipfile, numpy as np

with zipfile.ZipFile(path) as z:
    for member in z.namelist():
        if not member.endswith(".npz"):
            continue
        d = np.load(io.BytesIO(z.read(member)), allow_pickle=True)
        yield d                        # one sample at a time, without extracting to disk

3. Export to FBX / glTF / Blender

matrix_local + parents form a standard bone hierarchy; combined with vertices/faces/skinning weights you can build an Armature directly:

# Pseudocode: build the bone chain
for i, (name, parent) in enumerate(zip(d["joint_names"], d["parents"])):
    bone = arm.edit_bones.new(name)
    bone.head = d["joints"][i]
    bone.parent = None if parent < 0 else arm.edit_bones[d["joint_names"][parent]]

Then write W[i].indices / W[i].data for each vertex (keep the Top-4 weights per row and normalize) to finish skinning.

Notes and Limitations

• The Dataset Viewer being unavailable is expected behavior: the data is stored as .npz files inside zip archives, and Hugging Face's auto-detection heuristics can only parse formats such as Parquet / CSV / JSON / Arrow. It cannot read NumPy arrays inside archives, so the page shows "could not detect any supported data files". This is a format limitation and does not affect downloading or usage. If you do need an online preview, you can additionally upload a metadata.parquet (one row per sample with uuid / n_vertices / n_faces / n_joints / nnz) and point to it via configs.data_files in this card's YAML.
• Coordinates and units follow the original conventions of each source; before mixed training it is recommended to normalize consistently (e.g. scale to unit size by the bounding-box diagonal).
• Joints with body_joint_mask=False are mostly clothing/skirt/decorative bones; filter them as needed when learning skeleton topology.

License

This dataset is released under the TailorRig License (Copyright (C) 2026 Tencent; see the full terms in LICENSE at the repository root). Key constraints:

• Academic research use only; it must not be used for any commercial or production purposes;
• The original copyright and license notices must be retained upon redistribution or distribution;
• Provided "AS IS", without any express or implied warranty or liability.

Please confirm that your use case complies with this license before using the dataset.