started making a lora based on an art signed by "Prissy". very hard to tell if it's all made by the same artist or not...
the images i got are very low resolution and low quality so it doesn't work very well right now. it's probably going to need some synthetic dataset to fix the eyes or the best you can do is inpaint then with a lower weight
Description
trained on 16 images. most of the around 400 pixels on the longest side... lots of jpeg too.
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 0,
"networkDim": 16,
"numRepeats": 30,
"resolution": 512,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 8,
"noiseOffset": 0,
"targetSteps": 1067,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Adafactor",
"textEncoderLR": 0.0001,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 9,
"flipAugmentation": true,
"lrSchedulerNumCycles": 4
}