Honovy Style

训练参数如下。事后看来，我或许可以在每张图片的重复次数上更大胆一些。

281 张图片。

{

"engine": "kohya",

"unetLR": 1,

"clipSkip": 2,

"loraType": "lora",

"keepTokens": 1,

"networkDim": 16,

"numRepeats": 3,

"resolution": 1024,

"lrScheduler": "cosine",

"minSnrGamma": 5,

"noiseOffset": 0.1,

"targetSteps": 4215,

"enableBucket": true,

"networkAlpha": 8,

"optimizerType": "Prodigy",

"textEncoderLR": 1,

"maxTrainEpochs": 20,

"shuffleCaption": true,

"trainBatchSize": 4,

"flipAugmentation": false,

"lrSchedulerNumCycles": 3

}