//=time() ?>
@noop_noob test5 has slightly more training images than 1.3.0, and the output is close to 1.3.0, but degraded.
In 1.3.0, I did a 512x512 cropping step, but in test5 I did not do that, but used aspect ratio bucketing.
Maybe the same procedure with the exact same dataset would reproduce 1.3.0
@noop_noob Yes...
I am currently testing a WD1.4 based model that dataset with different contents to verify why v1.3.0's outputs are so nicely.
Even if I increase the number of training images, I cannot exceed the miraculously created 1.3.0. I cannot reproduce it... 😭