dataset: 26 photos tagged with realistic. most close-up or lower body. no navel/knees shown. lots of lowres upscaled with ai.
8/8 dim/alpha, adafactor 1040 steps/20 epoches, 0.0005 lr, cosine with 3 restarts. 5 snr gamma. no noise offset. shuffle tokens/flip images.
should have used less learning rate. 0.0001 was too low, 0.0005 was too high. maybe 3? also may be upscaling is bad. should have cropped more things out.
because i cropped so hard it doesn't mess the anime style as much at 1 weight but still does a bit. but lower weights lose details hard. i didn't have lots of good angles. it can do pov pretty well tho lol