a lora trained only on photos of solo sissies/femboys wearing chastity cages because i had too many photos of this concept and it's easier to train a chastity lora when all images are in the exact same angle
v1.1
main trigger: pov, 1boy, solo, otoko no ko
alt trigger: chastity cage focus, close-up
cage tags:
chastity cage (66)
chastity cage strap (16)
nub chastity cage (21)
flat chastity cage (12)
inverted chastity cage (2)
see-through designs are tagged with penis, uncensored or small penis, uncensored
color is metal by default
pink chastity cage (18)
black chastity cage (6)
blue chastity cage (1)
white chastity cage (1)
locations
on bed (47)
on couch (3)
on chair (3)
on floor, bathtub, partially submerged, water, foam (1)
on floor, beach, ocean, water, crowd, horizon, blue sky, cloud (1)
on floor, pool, water, fence, building (1)
poses
legs together (20)
knees up (17)
knees apart feet together (7)
squirming (6)
holding sex toy (5)
crossed ankles (4)
feet out of frame (4)
erection (3)
Description
trained on 81 photos. most photos weren't cropped at all. some images are chastity cage close-ups to train the detail better. the photos were tagged "photorealistic" but this lora probably can't be used to render drawings at all. none of the photos include people's faces for the obvious reason that the camera is pointed at their dick but the lora will still render faces more "realistic" if you try to use it with "1girl, looking at viewer"
there were a few photos with objects on the person's legs/thighs/crotch that i wanted the lora to learn but the lora failed to learn it. it's proly because i chose my tags poorly and tagged it on leg once, on thigh once and on crotch once. if i had tagged everything on body maybe it would have worked.
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 3,
"networkDim": 32,
"numRepeats": 20,
"resolution": 1024,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 0,
"noiseOffset": 0,
"targetSteps": 3600,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Adafactor",
"textEncoderLR": 0.00008,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 9,
"flipAugmentation": true,
"lrSchedulerNumCycles": 4
}