trained on 25 images. 18 photos. most images cropped to show only neck area. some images cropped to show only the ring or half of the ring. 2 portrait illustrations with face visible.
v1 was overtrained on neck area and was inconsistent so i removed images that didn't have the exact same design like spiked chokers and chokers with more studs. the design i kept had only 2 studs. i also used images that showed only half of the neck (like the left side only) to train the details better and avoid burning the neck composition. the portrait images used featured a character that is already in illustrious and an artist style that is also already in illustrious.
v1.1 was also trained on multicolored chokers. same design but left/right bands have different colors.
{
"engine": "kohya",
"unetLR": 0.0002,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 8,
"numRepeats": 16,
"resolution": 1024,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 5,
"noiseOffset": 0,
"targetSteps": 1200,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Adafactor",
"textEncoderLR": 0,
"maxTrainEpochs": 12,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": true,
"lrSchedulerNumCycles": 4
}
