Significant metric gap with appearance embeddings

[LaFeuilleMorte]

Hi, I've been testing the effect of the appearance embeddings on the metrics. With appearance_dim=32, my PSNR was 22 dB. While with appearance_dim=0, my PSNR=23.5dB. Which was a huge gap with this hyperparameter. Can you please tell me the effect of this parameter? Thank you so much!

[tongji-rkr]

Apperance Embbeding is designed to handle the problem of different exposures or different styles in the training images. When using it, set eval=False.