Linzhan's picture
Point at the open data_process pipeline that turns this raw layer into UniML3D
76acab9 verified
|
Raw
History Blame Contribute Delete
5.85 kB
---
license: other
license_name: mixed-per-object-see-notice
license_link: https://huggingface.co/datasets/Linzhan/Objaverse-XL-Rigged-Animated/blob/main/NOTICE.md
pretty_name: Objaverse-XL Rigged & Animated Renders
language:
- en
size_categories:
- 10K<n<100K
task_categories:
- text-to-3d
tags:
- 3d
- animation
- rigging
- skeletal-animation
- motion
- objaverse
- render
- multi-view
- video
---
# Objaverse-XL Rigged & Animated — Renders
Visual companion to
[**Linzhan/Objaverse-XL-Rigged-Animated**](https://huggingface.co/datasets/Linzhan/Objaverse-XL-Rigged-Animated),
which holds the 7,373 rigged-and-animated GLB assets themselves. This repository holds only what
was rendered from them: a four-view video of every animation clip, and a rest-pose grid per asset.
They live apart from the assets because they are bulky and numerous — 10,355 clip folders — while
the asset repo stays a compact 7,373 GLBs plus two tables. Nothing here is needed to use the
assets; it is for previewing motion on the Hub, for VLM captioning, and for eyeballing a rig
before you commit to processing it.
| | |
|---|---|
| clip renders | 10,355 folders · 41,420 MP4s (2.2 GB) |
| rest-pose grids | 7,373 PNGs |
| rendered frames | 957,818 per view (~8.9 h) · 3,831,272 across four views |
| resolution / rate | 512×512 · 30 fps |
## Contents
```
glb_render/ 10,355 dirs four-view render of every skeleton-driving clip
tpose/ 7,373 .png rest-pose render grid per asset
```
Stems match the asset repo exactly, so both join back to `metadata.csv` / `animations.csv` there
without any name mangling.
## Renders (`glb_render/`)
A four-view render of every clip that drives a skeleton and runs at least 5 frames — 10,355 of
the 16,190 rows in the asset repo's `animations.csv` (the rest animate only non-joint nodes, or
are too short). One folder per clip, named `{glb stem}-{action}`:
```
glb_render/00064e6f6eec4fd7acf27091b8de4d70-Object_4/
v000.mp4 v000.json front (azimuth -90°)
v001.mp4 v001.json back (azimuth +90°)
v002.mp4 v002.json left (azimuth 180°)
v003.mp4 v003.json right (azimuth 0°)
```
An asset with several actions contributes one folder per action; distinct actions that sanitize
to the same name get a `_00`/`_01` suffix in discovery order, so folder names stay unique.
Rendered with Blender EEVEE at 512×512, 30 fps, elevation 0, and composed into one H.264 MP4
per view — the transparent-film frames flattened onto a light-grey background, since H.264
carries no alpha. The MP4s preview each motion directly on the Hub; the per-frame RGBA PNGs
behind them are not hosted (they would add roughly 2.9M files), so extract stills from the MP4s
(`ffmpeg -i v000.mp4 frames/%04d.png`) if you need them.
Rendering covers the first 200 frames of a clip (~6.7 s); 2,150 of the 10,355 clips are longer
and are truncated — `frame_start`/`frame_end` in the per-view JSON record the covered range,
not the source range.
Each `v00X.json` holds the camera in NeRF convention: `view`, `azimuth`/`elevation` (radians),
`resolution`, `camera_angle_x` (horizontal FoV, radians), the 4×4 camera-to-world
`transform_matrix`, the `scene_scale`/`scene_offset` that normalised the asset into frame,
and `frame_start`/`frame_end`/`fps`.
## Rest poses (`tpose/`)
`tpose/<stem>.png` is a 1024×1024 2×2 grid of the asset's rest pose — front and back on the top
row, left and right below — rendered (EEVEE) from its GLB, one per asset in the asset repo's
`glb/`. Useful for classifying an asset's body plan without importing it.
These previously lived in the asset repo and were moved here so that everything rendered sits in
one place.
## Usage
```python
from huggingface_hub import snapshot_download
# just the rest-pose grids
snapshot_download("Linzhan/Objaverse-XL-Rigged-Animated-Renders", repo_type="dataset",
allow_patterns="tpose/*")
# one clip's four views
snapshot_download("Linzhan/Objaverse-XL-Rigged-Animated-Renders", repo_type="dataset",
allow_patterns="glb_render/00064e6f6eec4fd7acf27091b8de4d70-Object_4/*")
```
## Where these renders are used
The renders here feed the captioning and body-plan classification stages of the open pipeline at [`UniMate/data_process`](https://github.com/Friedrich-M/UniMate/tree/main/data_process), which turns the companion asset repository into the processed motion release [UniML3D](https://huggingface.co/datasets/Linzhan/UniML3D).
```bash
bash data_process/scripts/run_download.sh objaverse_renders
```
## Licensing and attribution
These renders **depict third-party assets** and are included as visual reference. They are not
covered by any blanket licence: each render inherits the upstream licence of the asset it shows,
exactly as that asset's GLB does. Resolve the licence for a given `object_id` through the
Objaverse-XL annotations before using or redistributing a render, and see
[`NOTICE.md`](https://huggingface.co/datasets/Linzhan/Objaverse-XL-Rigged-Animated/blob/main/NOTICE.md)
in the asset repo for the full statement and the takedown process.
Rights holders who want an asset removed can open an issue on either repository.
## Citation
This dataset is part of **UniML3D**, the training corpus introduced in
[UniMate: One Unified Model to Animate Diverse Skeletons](https://arxiv.org/abs/2609.05415)
(SIGGRAPH Asia 2026; [paper page](https://huggingface.co/papers/2609.05415)). If you use it, please cite the paper:
```bibtex
@article{mou2026unimate,
title = {UniMate: One Unified Model to Animate Diverse Skeletons},
author = {Mou, Linzhan and Lei, Jiahui and Dou, Zhiyang and Cai, Chenyue and Song, Chaoyue and Finkelstein, Adam and Rusinkiewicz, Szymon},
journal = {arXiv preprint arXiv:2609.05415},
year = {2026}
}
```