VfxDB

A large-scale OpenVDB volumetric-effects dataset: 11 categories, ~9k sequences, ~1M frames of smoke, fire, liquid, explosions, and related VFX volumes. Full payload is about 4.6 TB.

• 🌐 Project page: https://vfxdb-official.github.io/VfxDB/
• 📄 Paper: https://dl.acm.org/doi/10.1145/3799902.3811178
• 💻 Training code: https://github.com/VfxDB-Official/VfxDB
• 🧠 Checkpoints: https://huggingface.co/ryogishiki/VfxDB-models

License: CC BY-NC 4.0.

Categories

Category	Frames
BuoyantExplosion	240,000
CloudWave	10,463
EnvironmentalFog	23,843
IsotropicBurst	235,882
LiquidSplash	92,344
RingBlast	239,879
RisingFlame	10,158
SmokePlume	10,188
SurfaceFire	24,000
ViscousFlow	99,905
VortexColumn	50,250

Quick download

Do not clone this dataset repository. Install the Python deps, fetch the single entry script, then pull a starter slice.

python -m pip install "huggingface_hub[cli]>=0.36.0" rich zstandard

hf download ryogishiki/VfxDB tools/download_extract_data.py \
  --repo-type dataset \
  --local-dir .

python tools/download_extract_data.py /data/vfxdb --preset smoke

smoke downloads two sequence tars from every category. Expect /data/vfxdb/<Category>/category_index.json, per-sample JSON under index/, and a few extracted .vdb sequences. Not the full 4.6 TB.

The script pulls the rest of the downloader itself. If you already have the training repo, skip hf download and run python tools/download_extract_data.py /data/vfxdb --preset smoke from that checkout.

Want more?

# ~20% of all tars
python tools/download_extract_data.py /data/vfxdb --preset medium

# everything (~4.6 TB)
python tools/download_extract_data.py /data/vfxdb --preset full

# one or more categories, about N usable samples each
python tools/download_extract_data.py /data/vfxdb \
  --category CloudWave \
  --max-samples 1000

# interactive picker
python tools/download_extract_data.py /data/vfxdb --tui

• medium / full — same tree as smoke, just more sequences.
• --category + --max-samples — whole tars only, so you may get a bit over N.
• --tui — pick a mode in the terminal; same files as the CLI.

Pin a snapshot with --revision <commit> if you need a frozen copy.

Layout after extract

/data/vfxdb/
  SurfaceFire/
    category_index.json
    index/
      <sample>.json
    <sequence>/
      <sample>.vdb

Train and run the paper models from the training repo.

A small number of known-bad VDB files are skipped by default. Keep them with --include-bad. Downloader details: docs/DOWNLOADER_SPEC.md.

Citation

@inproceedings{VfxDB,
author = {Shu, Junwei and Liu, Hantang and Miao, Dawei and Song, Wenzheng and Yuan, Mingyang and Liu, Wenjie and Chen, Changgu and Li, Yang and Wang, Changbo},
title = {VfxDB: A Visual Effects Volume Dataset and Benchmark for VDB-Native Generative Modeling},
year = {2026},
isbn = {9798400725548},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3799902.3811178},
doi = {10.1145/3799902.3811178},
booktitle = {Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers},
articleno = {195},
numpages = {11},
keywords = {VDB, Visual Effects, Diffusion Generative Model, Machine Learning},
location = {
},
series = {SIGGRAPH Conference Papers '26}
}