numeral issue about the power over than 3

[BossunWang]

Hi experts,
The loss function use monomial to align distribution between stylized and style images.

cmd_styletransfer/losses.py

Line 53 in e398885

m_x = (x - c1_x).pow(i).mean(dim=self.axis)

However, assumed the feature values to smallest or highest, if some value pass with sigmoid function mapped between 0.0 and 1.0 which almost distribution is concentrated in 0.0 or 1.0!
And Then proposed method not useful for calculating loss over than pow 3 which closed to 0.0, but original feature distribution have "large oscillation".

You can see the results from other paper, the sky (low frequency) or fog(like noise) mapping to wrong style patterns.
https://arxiv.org/pdf/2203.07740.pdf