Shield:
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
Download (version 2): full dataset | sample dataset | single subject
Project Page | Dataset Page | Colab Quickstart
Cafca is a large-scale, diverse multiview face dataset for avatar reconstruction and animation, face parsing, synthetic-to-real transfer learning, animation, relighting, camera estimation, and more.
The full dataset contains 1,500 synthetic subjects, each of which is rendered in 3 environments with 13 random expressions from 30 views, resulting in 1.755 Mio. images. The dataset is split into 15 chunks. Each chunk is about 83 GB and contains 100 identities.
If you want to get started quickly, you can use the colab or download a subset of the dataset from the links above.
| Version | Date | Description | Author |
|---|---|---|---|
| 1.0 | 2024-09-13 | Initial release | Marcel C. Buehler |
| 1.1 | 2024-10-22 | Add in-the-wild dataset | Marcel C. Buehler |
| 2.0 | 2025-09-19 | Add FLAME annotations for first 100 subjects | Marcel C. Buehler |
| 2.1 | 2025-09-30 | Add FLAME annotations for all 1,500 subjects | Marcel C. Buehler |
After unzipping, the dataset is structured as follows:
SUBJECT/EXP_ENV/{cameras_json, color_image, foreground_mask, segmentation, 'landmark2d/STAR
SUBJECT and EXP are 5-digit numbers. ENV is a 3-digit number. Camera names go from C00 to C29. Image data (RGB, foreground masks, and segmentation maps) are saved as PNG; the cameras as JSON.
Each subject has folders with FLAME pseudo-GT annotations (flame_2023_ENV, e.g., flame_2023_000 for environment 000).
flame_2023_000
Please see the colab for an example on how to load a particular scene.
The dataset contains each expression rendered in three different environments. The first environment (index 000) is
the same for all expressions (Laval_Indoor_9C4A5690_8k.exr). The other two environments are picked at random from
the Laval Indoor Dataset. Environment indices are consistent within an identity but not across identities. The environment name is saved in environment.json in
the color_image folder.
The neutral expression (00000) is fixed for all samples. The expressions (00001...00012) are sampled randomly for each environment. Note that the expression indices are not consistent across different environments.
The dataset uses a right-handed coordinate system following OpenCV convention (X right, Y down, Z forward). For convenience, the camera json file contains redundant information:
the full projection matrices (P), extrinsic and intrinsic matrices (world2cam / cam2world and K), and all of
these parameters individually.
The segmentations are stored as grayscale 16-bit PNG. Region 0 is the background, and other entries refer to regions like facial skin, throat, hair, upper body, eyes, etc.
The dataset includes pseudo-GT annotations for keypoints and FLAME 2023. These annotations were obtained by running VHAP, a photometric fitting pipeline.
The annotations are stored as npz files that combine all expressions for a particular frame and environment. For example, the FLAME parameters for subject 00000 and environment 000 are stored under
00000-00099/00000/flame_2023_000/tracked_flame_params_100.npz.
2D Keypoints are estimated with STAR and stored as npz in the landmark2d/STAR subfolder for each frame and camera. For example, the following path stores landmarks for subject 00000, expression 00001, environment 000, and camera C00: 00000-00099/00000/00001_000/landmark2d/STAR/C00.npz
Please let us know how we can improve the documentation and help you get started.
If you find this dataset useful, please consider citing:
@incollection{buehler2024cafca,
title={Cafca: High-quality Novel View Synthesis of~ Expressive Faces from Casual Few-shot Captures},
author={Marcel C. Buehler and Gengyan Li and Erroll Wood and Leonhard Helminger and Xu Chen and Tanmay Shah and Daoye Wang and Stephan Garbin and Sergio Orts-Escolano and Otmar Hilliges and Dmitry Lagun and Jérémy Riviere and Paulo Gotardo and Thabo Beeler and Abhimitra Meka and Kripasindhu Sarkar},
year={2024},
booktitle={ACM SIGGRAPH Asia 2024 Conference Paper},
doi={10.1145/3680528.3687580},
url={https://doi.org/10.1145/3680528}
}
This dataset is not an official Google product. It is not supported by Google and Google specifically disclaims all warranties as to its quality, merchantability, or fitness for a particular purpose.