question about LPIPS loss

[avatarstones]

First thanks for the great ideas in the paper.

One question about LPIPS loss, in paper your said:
To capture fine details and further improve the realism,
we follow the Learned Perceptual Image Patch Similarity
(LPIPS) loss in [Zhang et al., 2018] and adversarial objective
in [Choi et al., 2020].

My understanding is LPIPS is image difference between 2 image, for instance
LPIPS(x,y)
In the loss function, this LPIPS is the difference between which 2 images?

Thanks.

[avatarstones]

Another question about cycle loss, in your paper

The cycle loss is a supplement of pixel supervision and can help generate high-fidelity results: L cyc = ||I t − G(I r , I t )|| 1 ,

Here you consider G(I_r, I_t) should be the same as I_t.
My question is, I_r has I_s face and I_t attributes, I_t of course has target face and attribute, then G(I_r, I_t) should have source face and target attribute, which should be similar to I_r, not I_t.
Should it be G(I_t, I_r), this will get an image of target face and target attribute, the result should be same as target I_t.

Thanks.