How can I evaluate GAN generated images quantitatively?

I am generating new images from a dataset of few thousand original images using DCGAN. I would like to evaluate how good my GAN performs. I am aware of the Inception Score (IS) and Frechet Inception Distance (FID). However, I am hesistant to use these since they utilize a pre-trained classification network whose classes are nowhere near my original images and thus will not be able to identify the new images. What quantitative measures should I use for my case?

Answer 1

[Raynier Suresh]

Average Log Likelihood, AM Score, Geometry Score, Precision and recall, Tournament based Method are some of other quantitative measures for GAN.

For Inception Score (IS) and Frechet Inception Distance (FID) instead of using a random pretrained network try using a custom network trained on real images.

For more info on training GAN refer the below link:

https://www.mathworks.com/help/deeplearning/ug/train-generative-adversarial-network.html#TrainGenerativeAdversarialNetworkGANExample-9