Everyone's favorite movie short king.
For the Anima version, I recommend starting at a weight of 0.8 if you want a clean, semi-realistic look. Full weight does work but dark lighting and reduced consistency starts creeping in.
The following tags were used to further separate the dataset into subsections. You can optionally use one of these tags if you want to nail a specific look.
Anima:
@gotg1- Orange outfit in first Guardians movie
@prison - Prison outfit in first Guardians movie
@gotg2 - Second Guardians movie
@gotg3 - Blue outfit in third Guardians movie
@casual - Civilian clothing in third Guardians movie
@infinitywar - Outfit in Infinity War
@endgame - Outfit in Endgame
Illustrious:
gotg1 - Orange outfit in first Guardians movie
pr1son - Prison outfit in first Guardians movie
gotg2 - Second Guardians movie
gotg3 - Blue outfit in third Guardians movie
c4sual - Civilian clothing in third Guardians movie
1infinity w4r - Outfit in Infinity War
endg4me - Outfit in Endgame
Description
Cleaned up the dataset. Added images from Infinity War and Endgame. Trained with a larger step count.
The training params are below. I'm not as keyed up on character loras as style loras so please don't assume these are the best params possible.
142 images:
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 16,
"numRepeats": 4,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 2130,
"enableBucket": true,
"networkAlpha": 8,
"optimizerType": "Prodigy",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 15,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}