How to implement Zero-Centerent Gradient Penaly with Keras

Question

I found a paper about text-to-image generation with GANs and in this article the authors show a loss function with a Zero Centered Gradient Penalty.

Gradient Penalty