This is the official repository following work on the 2019 NeurIPS workshop paper Approximating Human Judgment of Generated Image Quality, to approximate perceptual realism scores generated by HYPE from raw pixel data using neural networks.

The goal is to provide image-level labels of human perceptual realism, instead of distribution level, which most scores like FID, Inception Score, precision, HYPE, etc provide. i

Specifically, the model used was a DenseNet-121. Images used were CelebA and generated faces from StyleGAN, ProGAN, BEGAN, and WGAN-GP, trained on CelebA. The held-out validation and test sets were densely labeled by five human annotators using methodology similar to HYPE.

These scores are then compared with NVIDIA's image-level realism score, which is the most recent score released (though it was not evaluated against human judgment).

NB: A good portion of this code is from the Stanford ML Group's starter repository, which my collaborators and students in that lab and I wrote collaboratively. The remaining contributions for this project are mine.