SoroushMehraban/TopoRig_Dataset

TopoRig

Published dataset: SoroushMehraban/TopoRig_Dataset.

TopoRig contains neutral facial images, facial expressions labeled by action unit (AU), and rigged meshes in ICT and custom topologies for facial animation.

Dataset contents

Collection	Records	Contents
neutral_imgs	3,498	One neutral PNG per identity
expressed_imgs	100,331	Expressed PNGs for 3,466 identities, indexed by identity and AU
meshes_ict	3,497	Rigged FBX meshes in ICT topology
meshes_custom	3,497	Rigged FBX meshes in custom topology, with 3,496 landmark JSON files

All 103,829 images are 1024×1024 RGB PNGs. Meshes are stored once per identity and topology. The local dataset contains PNG, FBX and JSON files; downloadable releases can package these files as WebDataset tar shards.

File organization

README.md
metadata/
neutral_imgs/img_00006.png
expressed_imgs/img_00006-26.png
meshes_ict/img_00006.fbx
meshes_custom/img_00006.fbx
meshes_custom/img_00006.json

Match images and meshes using the complete identity string recorded in metadata/manifest.jsonl. Identify a mesh by both its collection and identity, since the two mesh collections can share the same filename.

The two-digit expression suffix identifies a TopoRig AU in 0..52. For example, 26 is jawOpen. Control IDs and names are listed in metadata/action_units.json and follow the TopoRig AU_NAME table.

Landmark JSON files map MediaPipe landmark IDs to vertex indices in the matching FBX mesh. The training recipes specify Z up, +Y front and normalization to an extent of 2.0 when loading meshes. Stored FBX coordinates remain in their source coordinate system.

Metadata and availability

• metadata/manifest.jsonl: relative file paths, identities, collections, AUs, asset types, byte sizes and SHA-256 checksums.
• metadata/layout.json: layout version, file counts and availability information.
• metadata/availability.jsonl: the available collections and AUs for each identity.
• metadata/action_units.json: AU IDs and names.
• metadata/bad_*.txt: sample exclusions and their documented reasons.
• metadata/splits.json: split availability and experiment selection settings.
• metadata/validation.json: integrity and structural validation results.
• metadata/provenance.json: available source information and distribution status.
• metadata/ict_split.csv, metadata/custom_split.csv: original mesh assignments.
• metadata/ict_landmarks/, metadata/ict_eval_landmarks/: 624 training and 350 benchmark mappings.
• metadata/evaluation_manifest.csv: original 350 ICT + 350 custom benchmark.
• metadata/reproduction.json: recovered-file checksums, coverage and validation limits.
• metadata/waft.json: exact external WAFT version and weight acquisition records.

Expression coverage varies by identity; 32 neutral identities have no expressed images. img_00100 has neither mesh, and img_02270 has no custom landmark JSON. ICT mappings for the recorded train/val and benchmark selections are included under metadata/. All nonexcluded training and validation identities and all 700 benchmark rows have the required meshes and mappings. Use the availability and exclusion metadata when selecting additional samples.

Use with TopoRig

Set TOPORIG_DATA_ROOT to the dataset directory and keep writable caches in a separate location using TOPORIG_CACHE_ROOT.

from dataset.mesh_dataset import MeshDataset

dataset = MeshDataset(
    data_root="/path/to/TopoRig_assets",
    mesh_source="custom",
    asset_cache_dir="/path/to/asset_cache",
    cache_dir="/path/to/tensor_cache",
    split="all",
    action_units=[26],
    blendshape_name_mapping="canonical",
    mesh_up_axis="z",
    mesh_front_axis="y",
    exclude_mesh_ids=["img_02270"],
)

from dataset.image_dataset import ImageDataset

dataset = ImageDataset(
    data_root="/path/to/TopoRig_assets",
    asset_cache_dir="/path/to/asset_cache",
    cache_dir="/path/to/tensor_cache",
    split="all",
    action_units=[8, 9, 10, 11, 20, 21, 22, 23],
    exclude_person_ids=["img_00100", "img_02270"],
    use_mediapipe_landmarks=True,
    require_mediapipe_landmarks=True,
    mesh_up_axis="z",
    mesh_front_axis="y",
)

These examples require the TopoRig code and its relevant runtime dependencies. The supplied reference_assets/ directory includes the demo meshes and oral assemblies. Blender/FBX support, MediaPipe, and WAFT source and weights are separate dependencies. Follow the TopoRig preparation instructions for derived landmarks, mesh tensors and supervision caches.

Downloadable shard format

A packaged release contains toporig_manifest.json and tar shards for each collection:

neutral_imgs/shard-00000.tar
expressed_imgs/shard-00000.tar
meshes_ict/shard-00000.tar
meshes_custom/shard-00000.tar

The release manifest records archive members, materialized file paths, identities, AUs, sizes and checksums. Each mesh record includes a landmarks.bin field with its JSON mapping, or empty bytes when unavailable; has_landmarks identifies which records include a mapping. Shards target 512 MiB while keeping each record together.

Install the TopoRig code and its hub dependencies, then download this release including the recovered metadata supplement:

python -m dataset.hf_assets \
  --repo-id SoroushMehraban/TopoRig_Dataset \
  --revision main \
  --local-dir /path/to/TopoRig_download

Use --components meshes_ict meshes_custom to download selected collections. Pass the downloaded directory as data_root in the examples above. TopoRig extracts selected files into a separate asset cache and verifies their checksums. Keep downloaded assets read-only and allocate separate space for extracted files, derived caches and outputs.

Splits and experiment reproduction

The shard split named data represents the complete inventory. Original train/val/test assignments and benchmark membership are provided in metadata/. After recorded exclusions, training selects 499 ICT and 2,796 custom heads; validation selects 63 ICT and 350 custom heads. The separate benchmark has 350 heads from each source. Preserve the supplied CSVs, AU settings and exclusions.

The original payload commit fd4f1f9f1996c139db8bf0ff85b3048f6e250d61 lacks this metadata supplement. Download the updated release and record its resolved commit for reproducibility. All 228 original shards are unchanged.

Data inputs for both training stages, validation and the benchmark are covered. Users also need an accessible TopoRig code release, Blender/MediaPipe and the external WAFT dependencies documented in REPRODUCING.md. Derived caches must be prepared; a complete GPU training/evaluation run has not been performed as part of this metadata audit.

Provenance and use terms

The dataset is free to use, modify, and share for noncommercial academic research, education, and teaching under the TopoRig Academic Use License 1.0, with attribution. Commercial use requires separate permission. Underlying third-party rights remain with their respective holders.

Cite TopoRig, the dataset repository URL, and the exact revision used. A paper citation and complete generation-source records are not included in this release. The original provenance audit is retained in metadata/provenance.json.

The checkpoint directory checkpoints_main_and_ablations_20260908 is excluded from the dataset release. See REPRODUCING.md in the packaged release for download, preparation, training, and evaluation commands.

Release package

This release contains 228 tar shards (112.03 GiB), plus metadata and reference assets. All 114,319 original payload files were checksum-verified while packaging. Shard counts by collection:

• neutral_imgs: 3 shards
• expressed_imgs: 73 shards
• meshes_ict: 48 shards
• meshes_custom: 104 shards

See metadata/packaging_validation.json for the build record and REPRODUCING.md for the complete training/evaluation procedure.