I attempted to train a model on a character not present in the dataset, specifically "Fenrir" from "Granblue Fantasy." The dataset was generated by extending original artwork using "gemini3.0proimage," resulting in some missing details.

fenrir

For the third iteration of the zimage LoRA model, it was found that using "prompt words" didn't work well for characters that weren't present in the model. Therefore, a prompt-based approach was adopted. See the trigger words for details. When using it, place the entire "trigger phrase" at the beginning and then continue writing the prompt afterwards.

Files with SHA256: 6a7b8e7d01b8c5656d2d49cc493d5a26221931d45564465864494c9dcd1e445e