Complete rework of dataset, adding movie screenshots, as well as more nsfw art pics. Trained on much higher step count to prevent anatomy under training.
Below are the training params. As said the steps are probably a lot higher than they needed to be, but I wanted to prevent issues with his eyes again like in V1.
149 images:
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 2,
"networkDim": 16,
"numRepeats": 4,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 2980,
"enableBucket": true,
"networkAlpha": 8,
"optimizerType": "Prodigy",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": true,
"lrSchedulerNumCycles": 3
}
