Use @rakupote as the trigger word to apply the style.

For captioning I followed the recommendations from Anima prompting guide.

For example:

masterpiece, 1girl, character name, series name, @rakupote, {tags}, {natural language description}

For this particular LoRA I had to use an edit model to preprocess the dataset. Then I generated stable tags using a recent model with good performance. Finally, during training I used compact structured captions with a stable style trigger, shuffled descriptive tags, and short natural-language descriptions made using a local LLM, helping preserve the artist style while keeping details like character, pose, outfit, and composition editable.