This is a port of Anijima, re-using the dataset generated as-is for a Krea2 training. Even if it was tag based, it works well with NLP. No trigger word. Just add the LoRA to get this style.

Enjoy 🥰

PS: Training done onsite, average of the three epochs (6-7-8) done with sd_mecha.

{
  "lr": 0.0001,
  "steps": 2060,
  "engine": "ai-toolkit",
  "epochs": 10,
  "batchSize": 1,
  "ecosystem": "krea2",
  "keepTokens": 0,
  "networkDim": 32,
  "resolution": 1024,
  "lrScheduler": "cosine",
  "minSnrGamma": null,
  "noiseOffset": null,
  "networkAlpha": 32,
  "optimizerType": "automagic",
  "shuffleTokens": false,
  "textEncoderLr": null,
  "flipAugmentation": false,
  "trainTextEncoder": false
}

PS2: Full disclosure, this model has been pushed to the Krea2 LoRA contest but i am not fishing for likes and reactions. Only publish creations and push the like button if you actually like this model, i am aiming for the participation badge only has i love badges 🤣