It's supposed to help you get some natural looking backgrounds, the dataset is made from 100 frames equally extracted across 10 yt videos.

It's not perfect, but i haven't tested it fully yet

This is an old trained model. more will come, this is nowhere near good , or even decent, but let that sink in. I think it may have been inproper training data lol, will experiment more. do whatever with it

here is config:

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 2,
  "networkDim": 32,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 3500,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "Adafactor",
  "textEncoderLR": 0.00005,
  "maxTrainEpochs": 14,
  "shuffleCaption": true,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}