V1.0a uses training sets from others, and uses better pictures, but fewer. Should be more sensitive to spikes.
Trained on Illustrious-XL.
valentine eisenberg:
Prompt :
masterpiece, best quality, amazing quality, v413nt1n3, Optional Prompt :
2024, 2023, Negative Prompt :
bad quality, worst quality, worst detail, sketch, censor, optional negative prompt :
monochrome, greyscale, signature, url, artis tname, username, text, English text,Its training task was set by me very early, and now I find it in the corner.
The images in this training set remain unprocessed except for resolution scaling, yet they underwent extensive essential manual annotation. Empirical evidence suggests this approach proves effective.
Description
{
"engine": "kohya",
"unetLR": 1,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 32,
"numRepeats": 3,
"resolution": 1984,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 4095,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Prodigy",
"textEncoderLR": 1,
"maxTrainEpochs": 35,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}