Measuring the Consistency and Diversity of 3D Face Generation

3D-aware GANs have shown their impressive power on 3D controlling for synthesized portraits. While the plausible facial reality is achieved, the inherent 3D properties of the generated results have actually not been well analyzed. One of the reasons is that the wildly-used metrics, such as Inception Score (IS) or Fréchet Inception Distance (FID), focus more on the perceptual features rather than explicit 3D clues. In this article, we propose two novel 3D metrics, which measure the face consistency and diversity on a 3D level, to compensate for IS or FID on GAN evaluation. With such metrics, the generated faces are systematically analyzed on a wide range of GANs, where we obtain reasonable but different conclusions from the wildly-used approaches. Inspired by the performance on the proposed metrics, we then discuss what is crucial for the GANs to get robust 3D properties, and what may bring uncertainty or inconsistency to the generating procedure. Finally, we propose two simple but efficient solutions to contribute to superior 3D synthesizing accuracy across different architectures, which further demonstrate the significance of the proposed metrics.