Gram loss magnitude is negligible during training

[sunset-clouds]

Hi authors,

While training the model, I noticed that the Gram loss remains extremely small throughout training. Even when I change the weight assigned to the Gram loss, its magnitude is still negligible compared to the other loss terms. I was wondering if you have observed a similar phenomenon during your experiments.

This makes me wonder whether the Gram loss actually contributes meaningfully during optimization. In practice, does the Gram loss significantly improve the reconstruction performance, or is its effect typically minimal?

Thanks for your help!

[yucornetto]

Thanks for the note! We added gram loss in an earlier revision and found it helps better performance in the earlier stage, we stick to this design later and did not investigate further.