RealDenseFace-Ava256

Multi-view images and FLAME registrations prepared from Ava-256 for RealDenseFace: Real-time Monocular 3D Face Reconstruction from Dense UV-space Priors.

Paper · Project · Code

The dataset contains 256 subjects, 15,345 frames, and 367,983 images sampled from Ava-256 sequences.

Split	Subjects	Images
Train	243	349,348
Validation	13	18,635

Data

• RGB images (667 × 1024 JPG) and semantic segmentation labels (PNG).
• Camera intrinsics, world-to-camera extrinsics, and image bounding boxes.
• Per-frame FLAME identity, expression, pose, translation, and vertex offsets.
• Registered mesh vertices and joint positions, shared across views of the same frame.

Registration and segmentation annotations are generated automatically and may contain errors. See the Ava-256 datasheet for source dataset characteristics and limitations.

Files

Images/<subject>.tar                  # RGB images and segmentation labels
registrations/<subject>.npz           # FLAME parameters, offsets, vertices, joints
metadata/cameras.json                 # subject -> camera -> K/R/t/width/height
metadata/frames.parquet               # frame index and sequence names
metadata/images.parquet               # image index, bounding boxes, archive paths
metadata/subjects/<subject>.parquet    # image index for one subject
splits/{train,validation}_subjects.txt
schema.json                          # field definitions and coordinate conventions
tools/                               # reading, reconstruction, visualization
checksums.sha256

Each tar contains images/<frame>/<camera>.jpg and segmentations/<frame>/<camera>.png. Metadata paths are relative to the dataset root. IDs are strings with leading zeros; registration_row identifies the frame's row in its subject NPZ. Hugging Face Dataset Viewer provides train/validation index tables.

Usage

Install numpy Pillow pyarrow, then read a sample directly from its tar archive:

import sys
sys.path.insert(0, "tools")
from load_sample import load_sample

sample = load_sample(".", "20230405--1635--AAN112", "056626", "400951")
rgb = sample["image"]
foreground = sample["segmentation"] > 0
vertices = sample["vertices_world"]

For a partial download, obtain the subject's tar, NPZ, metadata/subjects/<subject>.parquet, and metadata/cameras.json.

Key conventions:

• vertices_world and joints_world are in Ava world coordinates, in millimeters. FLAME parameters and offsets retain model units; reconstruction applies an outer scale of 1000 after pose and translation.
• v_offsets are added after expression deformation and before pose corrective blendshapes and skinning. Identity parameters are fitted per frame.
• Camera extrinsics use X_camera = R @ X_world + t; camera axes point right/down/forward. Update intrinsics when cropping or resizing images.
• Resize segmentation labels with nearest-neighbor interpolation. JPGs already use sRGB.

See schema.json for full definitions. Reconstruction and visualization use the standard FLAME 2023 model (flame2023.pkl), obtained separately from the official website. Install requirements.txt, then run:

python tools/reconstruct_flame.py --registration registrations/20230405--1635--AAN112.npz --flame-model /path/to/flame2023.pkl
python tools/visualize_sample.py --root . --subject 20230405--1635--AAN112 --frame 056626 --camera 400951 --flame-model /path/to/flame2023.pkl --output overlay.jpg

Output of the visualization command above, saved as overlay.jpg (left to right: RGB input, FLAME registration overlay, and foreground mask):

License

The original Ava-256 dataset was created by Meta Reality Labs' Codec Avatars Team under CC BY-NC 4.0. This processed dataset and our annotation contributions use the same license to the extent of our rights. See LICENSE and THIRD_PARTY_NOTICES.md.

Processing includes frame/view selection, image color conversion, FLAME fitting with vertex offsets, and repackaging. FLAME is subject to its own license; its model and topology must be obtained separately.

Citation

Please cite Ava-256, RealDenseFace, and FLAME when applicable:

@article{li2026realdenseface,
  title={RealDenseFace: Real-time Monocular 3D Face Reconstruction from Dense UV-space Priors},
  author={Li, Linzhou and Shao, Tianjia and Zhou, Kun},
  journal={arXiv preprint arXiv:2608.09238},
  year={2026}
}
@inproceedings{martinez2024codec,
  title={Codec Avatar Studio: Paired Human Captures for Complete, Driveable, and Generalizable Avatars},
  author={Martinez, Julieta and others},
  booktitle={Advances in Neural Information Processing Systems, Datasets and Benchmarks},
  year={2024}
}
@article{FLAME:SiggraphAsia2017,
  title={Learning a model of facial shape and expression from {4D} scans},
  author={Li, Tianye and Bolkart, Timo and Black, Michael J. and Li, Hao and Romero, Javier},
  journal={ACM Transactions on Graphics},
  volume={36}, number={6}, pages={194:1--194:17}, year={2017}
}

Questions and feedback