>>8696520
>like trying to use loss for quality control, too divorced from life to matter.
How is that so? SSIM and FID are computed on fully generated images in RGB space. SSIM probably isn't particularly good for content evaluation, but it's very cheap and you lose basically nothing by computing it. FID, on the other hand, uses features obtained from Inception V3. Which is not so useful for anime images, but it's possible to swap Inception V3 to using features from wd eva02 or camie taggers (camie was also trained to classify artists).
>And how would you even judge what epochs to test?
You don't judge, you test each one because all you lose is a little time.