this lora is trained mainly on front images

trained on the same dataset (111 images) as previous version with minor adjustments to captions and different trainer settings that are supposed to work better (no restarts, no SNR, 32 alpha, no text encoder). this version wasn't upscaled with the anime upscaler but with the realistic upscaler so it gens more "realistic" images.

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 2,
  "networkDim": 32,
  "numRepeats": 2,
  "resolution": 1024,
  "lrScheduler": "cosine",
  "minSnrGamma": 0,
  "noiseOffset": 0,
  "targetSteps": 1110,
  "enableBucket": true,
  "networkAlpha": 32,
  "optimizerType": "Adafactor",
  "textEncoderLR": 0,
  "maxTrainEpochs": 20,
  "shuffleCaption": true,
  "trainBatchSize": 4,
  "flipAugmentation": true,
  "lrSchedulerNumCycles": 1
}