Official VfxDB Dataset Repository

VfxDB is a large-scale OpenVDB volumetric-effects dataset with more than one million smoke, fire, liquid, explosion, and related simulation samples.

• 🌐 Project page: https://vfxdb-official.github.io/VfxDB/
• 💻 Training code: https://github.com/VfxDB-Official/VfxDB
• ⬇️ Official downloader source: tools/

Repository layout

dataset_index.json
<Category>/category_index.json
archives/<Category>/<Sequence>.tar
meta/vfxdb_meta.tar.zst
meta/io_bad_vdbs.jsonl
meta/io_bad_vdbs.meta.json
tools/download_extract_data.py
tools/vfxdb_downloader.py
tools/vfxdb_tui.py

The VDB payload is stored as one WebDataset tar archive per category/sequence. Each archive contains VDB members plus _sequence_manifest.json, which maps archive member keys back to their dataset-relative source paths.

meta/vfxdb_meta.tar.zst contains the per-sample JSON files referenced by all 11 <Category>/category_index.json files. The archive and indexes are published together and must always be read from the same repository revision.

Known IO-bad sample manifest

This revision contains a sanitized manifest for the completed full-dataset native IO scan:

• Manifest: meta/io_bad_vdbs.jsonl
• Integrity and provenance: meta/io_bad_vdbs.meta.json
• Known IO-bad samples: 17,416
• Affected sequence archives: 515
• Manifest SHA256: 1849aec7228ed06aa23f1ef6834d01290ee275618b7342b085771a4845056173

The JSONL is deterministic and contains only repository-relative information; no local filesystem or quarantine paths are published. Each row provides:

{
  "category": "CloudWave",
  "sequence": "0",
  "source_relpath": "CloudWave/0/s0000__n0091.vdb",
  "archive_path": "archives/CloudWave/0.tar",
  "sample_key": "s0000_n0091",
  "member_path": "s0000_n0091.vdb",
  "reason": "exception",
  "return_code": 2
}

Consumers should read the manifest from the same repository revision as the indexes and archives. The official downloader removes listed VDB and JSON files after extraction by default and records deleted_bad_io_sample: true on their rows in the installed category indexes. Use --include-bad only when the known-bad files are explicitly required.

Because samples are bundled into sequence tar files, filtering does not reduce the tar bytes transferred over the network. It prevents known bad VDB members from being extracted into the final dataset root; the tar objects themselves are unchanged.

Download the manifest directly with:

hf download ryogishiki/VfxDB \
  meta/io_bad_vdbs.jsonl \
  meta/io_bad_vdbs.meta.json \
  --repo-type dataset \
  --revision main \
  --local-dir vfxdb-download

Official downloader

The downloader and Rich TUI are distributed directly in this dataset repository and are versioned with its indexes and archive layout. No training repository checkout is required.

Fetch the small tool files, then install their three Python dependencies. No system zstd command is required:

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

hf download ryogishiki/VfxDB \
  requirements-downloader.txt \
  tools/download_extract_data.py \
  tools/download_extract_meta.py \
  tools/vfxdb_downloader.py \
  tools/vfxdb_tui.py \
  --repo-type dataset \
  --revision main \
  --local-dir VfxDB-tools

python -m pip install -r VfxDB-tools/requirements-downloader.txt

This command requests only the listed source files. It does not clone or download archives/, category indexes, or metadata payloads.

Start the interactive terminal workflow with:

python VfxDB-tools/tools/download_extract_data.py /data/vfxdb --tui

The TUI prepares every required local category index and per-sample JSON before showing the data modes. It then presents an exact whole-tar plan with space and sample totals, supports returning to change the selection, and uses Hugging Face's cache for interruption-safe reruns.

The same tool has a non-interactive CLI for scripts and batch jobs. A bare invocation installs all category indexes, internal IO-bad controls, and every referenced per-sample JSON. It does not download a VDB tar:

python VfxDB-tools/tools/download_extract_data.py /data/vfxdb

Choose exactly one data-selection mode when VDB files are wanted:

# Presets: Smoke selects 2 sequence tars per category; Medium selects 20%
# of all tars in fixed balanced order; Full selects every tar.
python VfxDB-tools/tools/download_extract_data.py /data/vfxdb --preset smoke
python VfxDB-tools/tools/download_extract_data.py /data/vfxdb --preset medium
python VfxDB-tools/tools/download_extract_data.py /data/vfxdb --preset full

# A percentage of all categories, balanced in fixed order and rounded up to
# complete tar files.
python VfxDB-tools/tools/download_extract_data.py /data/vfxdb --percentage 10

# One or more named categories, with the same usable-sample target applied to
# each category and rounded up to complete tar files.
python VfxDB-tools/tools/download_extract_data.py /data/vfxdb \
  --category CloudWave \
  --category SurfaceFire \
  --max-samples 1000

Every mode downloads and installs whole sequence tars. Known IO-bad files are removed by default without changing tar selection or quotas. To retain them:

python VfxDB-tools/tools/download_extract_data.py /data/vfxdb \
  --category CloudWave \
  --max-samples 1000 \
  --include-bad

For reproducible runs, pin the repository commit returned by the downloader:

python VfxDB-tools/tools/download_extract_data.py \
  /data/vfxdb \
  --revision <40-character-commit> \
  --preset smoke

--revision selects the dataset revision. If the tool files themselves must be reproduced exactly, use the same full commit hash in the preceding hf download --revision ... command. The normative behavior is documented in docs/DOWNLOADER_SPEC.md.

Current status

• Full sequence archive set uploaded
• Dataset and category indexes uploaded
• Non-VDB metadata archive uploaded
• Full-scan IO-bad manifest published
• Dataset-local downloader, Rich TUI, specification, and regression suite