v2 is much more stable and works well with other LoRAs. But it is more stylized than v1. v1 has certain qualities that make it nice to play with but it is a very large LoRA and can be tough to work with some models with it. Default colorset is a cyberpunk teal and pink, with a animation style. Prompts are very important still so be sure to play with the prompt.

Trained using D-Adaptation at 1.0 Unet and 0.5 Text Encoder. 200 training images. "a anime" Regularization images (does this help make the style more flexible?). Noise offset: -0.1. Straight booru captions, no trigger words.More steps possible in the future to further increase the quality and add in more ghostly figures.

Files with SHA256: 607a32da2476d186c9ea074799fe3db15da7acca09439149e4f200b4b7f00841