Not gonna mince words, best Fellatrix style lora in this whole site, spent a lot of time, uploading 3rd training, cleaned the dataset manually removing all text and speech bubbles, choosing a varied 300 image dataset, learned from first 2, 3rd works like a charm for me, actually impressed, I have made many loras but most are nothing special so I dont post it, the outputs look really good, first generation might look a little undetailed, but I use fooocus and just upscaling 1.5X with it good enough for the most part, maybe inpaint the face details at the most. Thats all the yapping, trained right here on Civitai. parameters below.{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 32,
"numRepeats": 1,
"resolution": 1024,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 5009,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Adafactor",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 67,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}
