Anchee Style

以下为训练参数。如果我要重做，我可能会降低网络维度（network dim）和alpha，以观察是否能改善此LoRA中出现的轻微伪影。

548张图像：

{

"engine": "kohya",

"unetLR": 0.0005,

"clipSkip": 2,

"loraType": "lora",

"keepTokens": 1,

"networkDim": 32,

"numRepeats": 2,

"resolution": 1024,

"lrScheduler": "cosine",

"minSnrGamma": 5,

"noiseOffset": 0.05,

"targetSteps": 6028,

"enableBucket": true,

"networkAlpha": 16,

"optimizerType": "Prodigy",

"textEncoderLR": 0.00005,

"maxTrainEpochs": 22,

"shuffleCaption": true,

"trainBatchSize": 4,

"flipAugmentation": false,

"lrSchedulerNumCycles": 3

}