Tearmoon Empire ティアムーン帝国物語 || Mia Luna Tearmoon / Anne Littstein / ...

The updated model must be used on used on webui version >= 1.7.0

Intermediate checkpoints https://huggingface.co/alea31415/random_stuff/tree/main/tearmoon

This model uses my updated pipeline https://civitai.com/articles/2383 which constructs the dataset automatically from anime.

---

Wildcard:

Mia Mia, frills, blue dress, puffy short sleeves, blue bow Mia, necktie, school uniform, collared shirt, pleated skirt Mia long hair Mia long hair, blue dress AnneT AnneT, maid headdress, maid apron SionT SionT,epaulettes,white gloves ChloeT ChloeT, school uniform, glasses, red necktie, collared shirt, white shirt Tiona Tiona, dress Keithwood Keithwood, black cape Elise Elise, glasses, yellow shirt, collarbone Ludwig Ludwig, glasses, ascot AbelT AbelT, epaulettes GainT GainT, military uniform, holding sword Rafina Rafina, white dress DionT DionT, armor, epaulettes DionT helmet DionT helmet, armor, cape, knight Liora Liora, maid apron, black dress, juliet sleeves, hair ornament, leaf hair ornament Cyril Cyril, blush, suspenders Lynsha Lynsha, juliet sleeves, hair bow Lambert Lambert, red bowtie screenshots booru

No balancing at character appearance level cause the failure in learning secondary appearance. For mia long hair it is better to use later checkpoints from hugging face.

This is another experimental run for OFT where I set unet lr to 1e-4, text encoder lr to 2e-5, and embedding lr to 1e-2. The higher learning rate might hurt flexibility.