I read few articles about Learned Perceptual Image Patch Similarity (LPIPS) metric.
In some articles it says that the lower is better. but some articles say that higher is better.
which one is correct??
In some cases, higher LPIPS are suitable for evaluating the diversity of images generated by GANs, but if we are generating images or comparing images to see how different the original images are from the ones to be compared, lower LPIPS are suitable for that as well.