dataset is 81 images. mostly photos tagged realistic cropped tightly around the cage. some drawings. some had belt straps other diy string ones. lots of 256 pixel lowres images
32/32 dim/alpha. epoch 12 of 18 epoches/1067 steps 4 batch size. prodigy, cosine. bucketing flipping shuffling. 5 snr gamma, 0 noise offset.
style didn't get realistic at 18 epoches this time. idk why. still distortions on fingers/eyes. 12 epoch was used cuz of a mistake in the dataset. highest res pics had the nub covering only the tip of the penis with foreskin around and that wasn't the look i was going for.
cropping too hard helped keeping the style ig but it lost context so bcuz of the straps it thinks thigh straps and chokers have a cage in the middle too. might help to zoom out so it can see where the legs end.
i added some product pics of the cage unworn. forgot to tag it as unworn but it's tagged no humans. might have helped with the details but it still gets the details wrong. not sure what i can do about that tbh. will try to avoid cropping images too small next time and stop at 512. it's going to hurt style a bit but at least it will stop messing the details? not such how much 32 dim is helping with that. i thought cosine would help. maybe constant is better
the pov gens seem to get the details on the lock better tho it's not perfect. not sure why. but ig you can flip a pov upside down and tag it from above to cheat