this lora is trained mainly on front images
trigger: chastity cage, padlock, testicles
extra: heart-shaped lock
colors: pink chastity cage, pink padlock, etc.
v1.7 (made with ai toolkit)
my chastity cage loras are all trained on civitai with kohya. for six months now, civitai added a kohya replacement called "ai toolkit". i wrote an article then saying it sucked https://civarchive.com/articles/23119/ai-toolkit-training-sucks-lmao in beta. then later it became the default and i said it still sucked https://civarchive.com/articles/26868/ai-toolkit-training-still-sucks-lmao. now KOHYA IS GETTING REMOVED AND THE AI TOOLKIT STILL SUCKS. the problems are literally the same problems it has had for 6 months:
no tags in metadata
lora name is random in a1111
quality is worse
costs twice as much buzz for 20 epoches
but the changelog https://civarchive.com/changelog?id=246 implies they improved it based on community feedback (idk whose, certainly not my feedback) and "it's now performing as well or better than Kohya in most cases". well, here is a lora trained with it. it cost the same amount of buzz as the same lora in kohya. it uses the same settings and the same dataset. the results are clearly not as detailed.
unless i figure out a way of making the ai toolkit work, i probably won't be able to make any more decent chastity cage loras. maybe it needs different settings. idk. but the impression i get is that it just sucks at learning detail in general. since every lora i train with it lacks detail where kohya would make the detail work just fine.
v1.7 (more real)
trained on same dataset as v1.5 but with adjustments that should have improved quality. doesn't use anime upscaler so images will look more "realistic"
v1.5 (more anime)
not as good as v1.7 but doesn't make stuff realistic
Description
trained on 111 images incl 99 photos tagged "photorealistic" some colored drawings and 5 greyscale manga. all images were tagged with "chastity cage, padlock" and then additional features like "testicles" (102) and "pink chastity cage" (82) and "heart-shaped lock" (8). cropped images to better learn detail, 62 crotch focus and 20 "chastity focus". all images were 4x upscaled up to 1024 with anime R ESRGAN Anime6B to get rid of jpeg artifacts. some were pretty low res tbh and that might have messed with details like on the padlock
this version was mainly trained on front images because that's where the padlock is duh. there's some data from behind/side/on back but it's not very good at those
i had some pics of dildos/anal sex to it can gen penises normally but it's going to add padlocks to chokers.
i used close ups of photos to make the ai learn the details then drawings of characters wearing chastity with their heads visible so the ai doesn't get stuck rendering only the cage no matter what you prompt and to stop making the base style realistic/smooth. worked ok ig
i trained both a 32/16 dim/alpha and an 8/4 with the same amount of steps but the 8/4 ended up too weak. this is the 32/16. maybe if i had made it 8/16 would have turned better?
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 32,
"numRepeats": 2,
"resolution": 1024,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 5,
"noiseOffset": 0,
"targetSteps": 1110,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Adafactor",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": true,
"lrSchedulerNumCycles": 3
}FAQ
Comments (1)
YES THE PADLOCK THE MOST IMPORTANT PART.