Reference-based GAN Evaluation by Adaptive Inversion

Jianbo Wang, Heliang Zheng, Toshihiko Yamasaki; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 910-918

Common evaluation metrics for generative models such as the Frechet Inception Distance (FID) can produce variable results due to their reliance on image sampling. Some methods accelerate the convergence to optimal FID scores by utilizing pre-trained models for constructing discriminators a process that bears similarity to the operational mechanism of the FID metric itself. This approach can lead to an overestimation of performance. Consequently while the FID scores may improve the visual quality of the generated images may deteriorate. To better evaluate the visual quality of a GAN model we propose a reference-based evaluation metric for GAN by leveraging GAN Inversion. Specifically our evaluation metric measures the invertibility of a GAN model to invert unseen images from the target distribution. Experimental results demonstrate that our proposed evaluation metric could effectively measure GAN models under the same image content especially for those trained with pretrained vision models.