China's 200,000-scan 3D facial database aims to train more accurate digital human models

Digital humans - the photorealistic virtual faces that appear in film visual effects, video game characters, medical simulations, and increasingly in virtual communication - depend on accurate three-dimensional facial models. Getting those models right requires training data: large quantities of precisely measured, high-resolution scans of real human faces across a wide range of ages, expressions, and ethnicities.

Most existing facial databases are too small, too reliant on 2D photography, or too dominated by a narrow demographic range to train models that perform well across diverse populations. A research team at the Shenzhen Institutes of Advanced Technology, part of the Chinese Academy of Sciences, built something considerably larger.

The database they assembled contains approximately 200,000 high-fidelity 3D facial scans, making it the largest high-precision 3D facial dataset built in China. It was selected for inclusion in Fujian Province's 2025 High-Quality AI Dataset Program. But the database itself is only part of what the team produced.

The landmark problem in 3D face modeling

A core technical challenge in facial modeling is landmark detection: identifying specific anatomical reference points on a face - the corners of the eyes, the tip of the nose, the contours of the lips - precisely enough to anchor a three-dimensional model. These landmarks serve as the skeleton on which everything else is built. Errors in landmark placement propagate through the entire model.

Most existing algorithms for facial landmark detection rely on 2D texture - the color and brightness patterns in a photographic image - either as a primary input or as an assist to 3D geometric data. This works reasonably well in controlled conditions but degrades under varying lighting, makeup, occlusion, or when the algorithm encounters facial characteristics that were underrepresented in its training data.

Some approaches circumvent the problem by using synthetic faces or template models as references. These can improve consistency but introduce their own limitations: real human faces do not conform perfectly to templates, and synthetic training data can introduce biases that show up as systematic errors in real-world application.

A network built on curvature

Led by Prof. Song Zhan, the team developed a curvature-fused graph attention network, abbreviated CF-GAT, that takes a different approach. Rather than relying on 2D texture or template models, the network works directly from the three-dimensional geometric structure of the face - specifically, the curvature of the surface at each point.

Curvature is a geometric property that describes how sharply a surface bends. Around the nose, eye sockets, and mouth, curvature changes in ways that are highly consistent across different faces and highly distinctive between different facial regions. These curvature patterns are largely independent of lighting conditions, skin tone, or photographic quality - which makes them a more robust signal than texture for identifying anatomical landmarks.

The CF-GAT architecture fuses curvature information into a graph attention network, a type of neural network that models relationships between connected points - in this case, the three-dimensional point cloud that a facial scan produces. By attending to both local geometry and the relationships between nearby surface points, the network predicts 3D landmark coordinates without requiring the 2D texture cues that previous methods depended on.

Testing showed improved robustness

Validation testing demonstrated that CF-GAT performed better than previous methods in conditions that challenge texture-based approaches: noisy scans, faces with limited prior representation in training data, and geometric variations that template-based methods handle poorly. The researchers describe improved generalization across diverse facial characteristics - meaning the network's performance held up more consistently when applied to faces that differed significantly from those in its training set.

This is practically meaningful for applications like medical imaging, where facial diversity and scan quality variation are both real constraints, and for entertainment and virtual communication applications where the goal is to accurately represent any human face rather than a statistical average.

Privacy and data ethics in facial databases

The construction of a 200,000-scan facial database raises questions that the press release does not address in detail: how subjects were recruited, what consent procedures were used, how the data is stored and accessed, and what restrictions govern its use. These questions are not peripheral. Large facial databases have been misused in the past - used to train surveillance systems, sold without adequate subject consent, or applied to purposes that subjects were not informed about when they participated.

The database's inclusion in a government-sponsored high-quality dataset program suggests it meets certain institutional standards, but the specifics of its data governance are not publicly detailed in the available materials. For researchers or institutions considering use of this data, those details matter.

The technical contribution - a curvature-based network that reduces dependence on 2D texture in 3D facial landmark detection - stands on its own merits and is separable from the database construction questions. Both aspects of the work will likely attract attention from the digital human research community.