mq_blend_pony版本v1.0 (ID: 708041)

V3:

Training data can be downloaded here (11 pictures).

Parameters:

{

  "engine": "kohya",

  "unetLR": 1,

  "clipSkip": 1,

  "loraType": "lora",

  "keepTokens": 1,

  "networkDim": 32,

  "numRepeats": 10,

  "resolution": 1024,

  "lrScheduler": "cosine",

  "minSnrGamma": 5,

  "noiseOffset": 0.1,

  "targetSteps": 1027,

  "enableBucket": true,

  "networkAlpha": 16,

  "optimizerType": "Prodigy",

  "textEncoderLR": 1,

  "maxTrainEpochs": 28,

  "shuffleCaption": true,

  "trainBatchSize": 3,

  "flipAugmentation": false,

  "lrSchedulerNumCycles": 3

}

At epoch #17, the results are best, though the pictures are distorted slightly ?. So you may need to decrease the lora strength (0.9 ~ 0.8).

This lora generates a girl with round face.

描述:

训练词语: mq_blend,1girl