Use @belmina as the trigger word to apply the style.For captioning I followed the recommendations from the <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima#prompting">Anima</a> prompting guide.For example:<pre><code>masterpiece, 1girl, character name, series name, @belmina, {tags}, {natural language description}</code></pre>I preprocessed and inpainted the dataset with an edit model, then generated stable tags with a community model. Training used compact structured captions, the @belmina style trigger, shuffled descriptive tags, and short local-LLM descriptions to preserve the style while keeping character, pose, outfit, and composition editable.

Files with SHA256: 7d3ac75d0d98548963b41ad404be11c5af2b256090a2aade7b04f89a9ee5b859