SynthFace Dataset Overview

Updated 28 January 2026

SynthFace is a synthetic face dataset created using parametric 3D models, GANs, and diffusion pipelines to offer controlled variations in identity, pose, and illumination.
The dataset consists of one million 96×96 images across 10,000 unique identities, ensuring detailed annotations and traceable parameters for each facial image.
Evaluations demonstrate that pre-training on SYN-1M enhances real-world face recognition accuracy, supporting research in privacy-preserving and bias audit applications.

SynthFace Dataset

SynthFace refers to a family of synthetic face datasets generated by parametric 3D models, generative adversarial networks (GANs), or diffusion-based pipelines, providing large-scale, richly annotated, and highly controlled facial images. Prototype datasets carrying the SynthFace name have been released across projects focused on deep facial recognition training, biometric fairness analysis, face landmark detection, and 3D face regression. The most prominent and widely cited variant is SYN-1M, introduced for reducing dependency on large real-world face collections by providing one million photorealistic images with continuous control over identity, pose, illumination, and background. Modern SynthFace frameworks have become a central tool in privacy-preserving research, domain generalization studies, and bias audit pipelines (Kortylewski et al., 2018).

1. Parametric Data Generation Methodology

At the core of the original SYN-1M (SynthFace) dataset is the Basel Face Model 2017 (BFM17), a 3D morphable face model representing facial shape and albedo in high-dimensional PCA spaces:

Shape synthesis: $S(\alpha) = \bar{S} + U_s\Sigma_s\alpha$ , with $\alpha \sim \mathcal{N}(0, I)$
Texture (albedo) synthesis: $T(\beta) = \bar{T} + U_t\Sigma_t\beta$ , with $\beta \sim \mathcal{N}(0, I)$
Expression variation: PCA-based subspace, sampled as Gaussian coefficients.

Illumination is modeled using Lambertian reflectance under a 27-coefficient spherical harmonics environment, with parameters $L \sim \mathcal{N}(\mu_L, \Sigma_L)$ empirically fit from the Basel Illumination Prior (BIP), itself derived from 14,348 real face images.

Camera and pose use a pinhole model, fixed focal length, and uniformly sampled Euler angles:

yaw ( $\theta_y$ ) $\sim$ Uniform $(-90^\circ, 90^\circ)$
pitch ( $\theta_p$ ) $\sim$ Uniform $(-30^\circ, 30^\circ)$
roll ( $\theta_r$ ) $\sim$ Uniform $(-15^\circ, 15^\circ)$

Backgrounds are randomly drawn from the Describable Texture Database (DTD).

Rendering is executed via off-screen OpenGL rasterization at $96 \times 96$ RGB, with no additional geometric or photometric augmentation (Kortylewski et al., 2018).

2. Dataset Composition and Properties

SYN-1M comprises:

10,000 unique identities (independent draws in shape and texture space)
100 images per identity covering the full range of pose, illumination, and backgrounds
1,000,000 total images
Image format: $96 \times 96$ px, 8-bit PNG, RGB
Illumination: 27-D BIP spherical harmonics
Backgrounds: $\sim$ 500 DTD textures

No additional data augmentations or synthetic perturbations are applied; this guarantees that all appearance and covariate variation is explicitly parameterized and traceable (Kortylewski et al., 2018).

3. Dataset Preparation, Usage Protocols, and Public Resources

Preprocessing uses MTCNN for face detection and alignment, resizing to $96 \times 96$ px. No augmentation (mirroring, color jitter, etc.) is performed within the SYN-1M pipeline.

Splitting scheme: All of SYN-1M is used for pre-training ("virtual domain") in the reference studies, while downstream validation and benchmarking are performed on real datasets (e.g., LFW, IJB-A, CMU-Multipie). For pipeline customization, users can hold out $5\%$ – $10\%$ for hyperparameter tuning.

Access: The complete synthetic face image generator, including parametric code, experiment scripts, and pre-trained models, is available via public repositories (https://github.com/unibas-gravis/parametric-face-image-generator and https://github.com/unibas-gravis/scalismo-faces). Adopters specify desired sample counts, pose/illumination ranges, and render faces directly ready for standard deep recognition pipelines (Kortylewski et al., 2018).

4. Experimental Evaluation and Benchmark Results

A. Synthetic-Only Training:

On CMU-Multipie, FaceNet-Synth achieves accuracy $\approx 0.89$ .
On LFW ("in-the-wild") accuracy is $\approx 0.80$ .
On IJB-A (TAR@FAR=0.1), performance is approximately $0.625$.

B. Real + Synthetic Data Strategies:

Pre-training on SYN-1M, then fine-tuning with 100K real images from Casia WebFace, LFW accuracy rises to $\approx 0.936$ (relative to a real-only baseline of $\approx 0.891$ ).
IJB-A TAR@FAR=0.1: $0.850$ (100K real), $0.882$ (200K real), $0.906$ (full 453K real).
On CMU-Multipie: accuracy $0.913$ (100K real), $0.933$ (453K real).

C. Observed Effects:

The real-to-virtual performance gap is pronounced for difficult benchmarks (LFW, IJB-A) but can be closed via even modest real-data fine-tuning ( $\sim$ 100–200K images).
No negative transfer from pre-training on synthetic data; all models benefit, outperforming real-only analogues.

These findings empirically demonstrate that synthetic pre-training with SYN-1M, even at a scale and image quality below current photorealistic GANs, yields substantial gains in cross-domain recognition robustness and data efficiency (Kortylewski et al., 2018).

5. Comparison and Influence on Modern Synthetic Face Pipelines

The SYN-1M dataset and generation protocol underlie many subsequent synthetic face frameworks:

Parametric 3D pipelines in later works (e.g., Baltrušaitis et al.) inherit explicit control and annotation strategies for geometry, lighting, and hair (Baltrusaitis et al., 2020).
Many GAN-based datasets such as SFace and SymFace use the "all identities synthetic, no real data" approach but substitute StyleGAN2-ADA for parametric models, achieving higher visual realism and privacy guarantees (Boutros et al., 2022, Grimmer et al., 2021).
Differentiable generation pipelines have increased photorealism and annotation richness; recent diffusion-based pipelines condition generative models on parametric depth and geometry maps, yielding ground-truth mesh and attribute annotation suitable for 3D face regression (Rowan et al., 2023).

The explicit metric-driven evaluation of SYN-1M results, dataset size, and attributes form a reference baseline and motivation for subsequent developments in synthetic data for facial analysis.

6. Use Cases, Limitations, and Future Directions

Applications:

Pre-training and fine-tuning face recognition architectures where privacy, legal, or demographic constraints limit the acquisition of real data.
Serving as a "virtual domain" for domain adaptation and transfer learning studies.
Benchmarking data efficiency and model robustness under controlled covariate distribution shifts.

Limitations:

Residual domain gap persists on unconstrained, in-the-wild real benchmarks; synthetic-only models underperform real-only counterparts unless fine-tuned.
Rendered image resolution ( $96 \times 96$ ) and controlled backgrounds are less realistic compared to state-of-the-art GAN/diffusion-based synthesis.
No direct modeling of fine-grained, high-frequency details (e.g., subtle skin texture, fine hair, occlusions).

A plausible implication is that increased photorealism and diversity in synthetic face datasets, via modern GAN or diffusion models, can further narrow the synthetic–real gap, especially if combined with parameterized annotation strategies initiated in SYN-1M. Scaling up to millions of identities, high-resolution images, and including rare or composite demographic categories are active directions.

7. Summary Table: SYN-1M (SynthFace) Core Properties

Property	Value/Description	Source
Model	Basel Face Model 2017 (BFM17)	(Kortylewski et al., 2018)
# Unique identities	10,000	(Kortylewski et al., 2018)
Images per identity	100	(Kortylewski et al., 2018)
Total images	1,000,000	(Kortylewski et al., 2018)
Resolution	$96 \times 96$ RGB, PNG	(Kortylewski et al., 2018)
Pose (yaw/pitch/roll)	Uniform in $[-90^\circ,90^\circ]/[-30^\circ,30^\circ]/[-15^\circ,15^\circ]$	(Kortylewski et al., 2018)
Illumination	27-D spherical harmonics (BIP)	(Kortylewski et al., 2018)
Background	$\sim$ 500 DTD crops	(Kortylewski et al., 2018)
Annotations	Perfect label via model params	(Kortylewski et al., 2018)
Public code	parametric-face-image-generator	(Kortylewski et al., 2018)

SynthFace (SYN-1M) remains a foundational resource in synthetic data research for facial analytics, providing controlled, rich, and diverse data for training, evaluation, and systematic analysis in deep face recognition and beyond.