Use @ouru as the trigger word to apply the style. Might not always be consistent.Recommended weight range is 0.95 - 1.15 depending on other settings.I added two variants:<ul><li>base version, no step count in name, all images in the gallery use it;</li><li>more train steps, a bit overtrained, in case you need stronger style;</li></ul>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, @ouru, {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 @ouru style trigger, shuffled descriptive tags, and short local-LLM descriptions to preserve the stylewhile keeping character, pose, outfit, and composition editable.

Files with SHA256: 787c2a38bc4b8daf93de30f1d3bbb39f31ee5c93cbc04b401031cb99cd1fdfea