<h2>Kyouyama Kazusa (Blue Archive)</h2>杏山カズサ (ブルーアーカイブ) / 쿄야마 카즈사 (블루 아카이브) / 杏山和纱 (碧蓝档案)<h3>Usage</h3><ul><li>Use any or all of the following tags to summon Kazusa: <code>kazusa, 1girl, animal ears, colored inner hair, halo, short hair</code></li></ul><ul><li>For her normal outfit: <code>black choker, black jacket, black pantyhose, green sailor collar, hooded jacket, white skirt, miniskirt, hairclip, pink neckerchief, school uniform, sneakers</code><ul><li>For a closed jacket, you might need to add <code>open jacket, white shirt</code> to the negative prompt.</li><li>For raised hood: <code>hood up</code></li></ul></li><li>For her "trying to look cool" expression: <code>expressionless, blush, sweatdrop, v-shaped eyebrows, embarrassed</code><ul><li>For selfies: <code>selfie, (reaching towards viewer:1.2)</code></li></ul></li><li>For eating: <code>eating, :t, holding food, food on face</code></li><li>Accessories: <code>macaron, cake, fork</code></li></ul>Not all tags may be necessary. Kazusa has a lot of high quality art so I was able to curate a very large dataset. As a result it should be easy to change her outfit and hairstyle with different prompts.I try to avoid overfitting my LoRAs. As a result, your prompt might require more tags to perfectly reproduce the character, but in exchange the model should be much more flexible when changing clothing or hairstyles and mixes well with other LoRAs and art styles. I recommend setting weight to 1.0 and using some of the tags provided above.私はLoRAモデルのオーバーフィッティングを避けるようにしています。そのため、キャラクターを完全に再現するには、より多くのタグが必要になる場合がありますが、その代わり、モデルはキャラクターの衣服や髪型の変更に柔軟に対応でき、他のLoRAモデルやアートスタイルとも相性が良くなります。必要に応じて、重みを1.0に設定し、上記で提供したタグのいくつかを使用することをお勧めします。저는 LoRA 모델을 오버피팅하는 것을 피하기 위해 노력합니다. 그 결과로, 캐릭터를 완벽하게 재현하기 위해서는 더 많은 태그가 필요할 수 있지만, 대신 모델은 캐릭터의 의상이나 머리 스타일을 쉽게 변경할 수 있으며, 다른 LoRA 모델 및 아트 스타일과도 잘 어울립니다. 가능한 태그 중 일부를 사용하고, 가중치를 1.0으로 설정하는 것을 권장합니다.我尽量避免LoRA模型的过拟合。因此，你的提示可能需要更多的标签来完美地复现角色，但是相应地，该模型在更改角色服装或发型方面更加灵活，与其他LoRA模型和艺术风格混合得更好。我建议将权重设置为1.0，并使用上面提供的标签中的一些。Refer to <a target="_blank" rel="ugc" href="https://huggingface.co/khanon/lora-training">https://huggingface.co/khanon/lora-training</a> for other Blue Archive characters, training parameters, and datasets.

Kazusa / カズサ / 카즈사 / 和纱 (Blue Archive)

Initial release.Since Kazusa had a very large dataset, I experimented with training for many more steps (almost 2x) and then resizing to a lower dimension than usual (16 instead of 32). It seems the larger dataset allowed me to train for much longer without overtraining.

Files with SHA256: 54ab8df42e151b882ec674b2127e7040c769f7bf99fc0d58c7c172cacfeacb5b