Type | |
Stats | 313 229 |
Reviews | (70) |
Published | Jan 15, 2024 |
Base Model | |
Training | Steps: 2,200 Epochs: 10 |
Usage Tips | Clip Skip: 1 |
Trigger Words | bayleef |
Hash | AutoV2 0B235473F9 |
Bayleef LoRA trained on NovelAI with Kohya training scripts using the DAdaptAdam optimizer at 512x512 resolution. Works quite well, but has a few issues. Skin coloration can be off depending on the base model used, YiffyMix 3 wanted to turn her skin green occasionally. Increasing LoRA strength can help, but may cause other issues to come out. Adding (yellow skin) to your prompt can also address this if you don't want to increase LoRA strength. Has a hard time with the 'neck pod' placement in certain poses. They were all over the place in the training data, which ended up being reflected in the final model. Tail placement can be unstable. Head leaf can come out mangled sometimes. Should have wide compatibility with model mixes based on NovelAI, which is most of them. Trained on a mixed male/female dataset.