trained on 39 low quality images without upscaling cropped as tight as possible around the cage. most 256 pixel and from videos/pics by lockedupslut. they were tagged source-lus1, 2, 3, 4 in the dataset. all tagged realistic. most tagged close-up. 3 pics cropped around the lock without showing the whole tube
4/4 dim/alpha. 18 epoches 1188 steps 4 batch size. adafactor 0.0003 LR cosine with 3 restarts. 5 snr gamma. no noise offset.
sucks but i only had lowres video screencaps to work with like 240p so not much i could do. no on_side pics either so i made one rotating the legs_up i had.
i thought 0.0003 lr would work based on prev experiments but before i had upscaled images so ig bucketing scaled down the lr of the lowres pics and since i only had lowres pics the real lr was much less so some gens don't even connect the tube to the ring right. so pro tip upscale the good pics downscale the shit pics if you train on civitai maybe that works
cropping kept realism style effect low but because i cropped too much it makes dildos metal too and keeps adding metal butt plugs without prompt. maybe some pics need to show the whole ass or a whole dildo in training to stop the lora from messing them up