dataset: 224 images / 50 multiple views, 17 anime screenshots, 10 greyscale, 5 sketches. upscaled to get rid of jpeg artifacts.
tagged with "anime screenshot" by mistake. should have been "anime screencap" but danbooru changed the tag name since illustrious was trained and i didn't realize it until later
no trigger word for base style
turned out pretty good i think after 4 failures and thousands of buzz wasted but could be better.
hands got messed up probably because some gloves that are fingerless for 1 finger.
tendency to give "hand up" because i forgot to tag it in all pics that had it.
missed some hair style tags too
wasn't sure how much sketch/screencap/greyscale was going to mess with the style. probably should have added more
lots of details i wanted to capture were ignored as usual. nipple piercings, navel piercings, earrings, chokers, o-rings. should probably add them as crops.
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 0,
"networkDim": 32,
"numRepeats": 3,
"resolution": 1024,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 5,
"noiseOffset": 0,
"targetSteps": 3360,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Adafactor",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}
