Model Description

This LoRA model is a enhancement model for Sparxie from Honkai: Star Rail, fine-tuned on a dataset of 300+ images. This LoRA can recreate the character using the trigger word alone, while also providing detachable clothing tags for creative flexibility.

In addition to the version based on Illustrious XL, I have also trained versions based on NoobAI, AnimagineXL, PonyXL, WAI-IllustriousXL, and ZukiAnimeIll, which will be released soon.

Recommended Settings

• Recommended weight: 0.8 - 1.0

• (Please adjust the prompts according to the official documentation of the checkpoint model you are using.)

• example prompt :

masterpiece, best quality, newest, absurdres, highres, Sparxie, 1girl, earrings, smile, looking at viewer, choker, earrings, gloves, white background, smile, cute, playful, hand on hip, simple background

For full-body shots, you can try adding full body and boots.

Usage Tips

• Clothing & Accessories: Special optimization has been applied to address the "hard-baked" clothing issue. Sparxie's official outfit is now detachable. Her official design is primarily bound to the following tags: hat, dress, choker, gloves, boots, earrings.

• Changing Outfits: In my testing with common checkpoint models, removing the official clothing tags allowed for adding any preferred clothing elements. If residual elements persist when using your specific checkpoint for inference, consider adding the aforementioned tags to the negative prompt.

Others

While I am very satisfied with the accuracy of Sparxie's features—including her facial mark**, mini hat, and earrings**—I cannot guarantee that every detail is perfectly stable yet. I plan to release an optimized version for better details soon.

If you have any feedback, please feel free to leave a comment.

Thank you for your support!

模型简介

这个LoRA模型是角色火花（Sparxie）的角色增强模型，在300+ 张图片的小数据集微调得到。此LoRA使用触发词即可还原人物，也提供了可拆卸的服装提示词，方便各类二创发挥。

除了基于Illustrious XL的版本，我还训练了基于NoobAI，AnimagineXL，PonyXL，WAI-IllustriousXL，ZukiAnimeIll的版本，稍后会发布。

推荐设置

• 推荐权重： 0.8 - 1.0

• （请根据你实际使用的Checkpoint模型的官方文档，适当调整提示词策略。）

• 示例prompt:

masterpiece, best quality, newest, absurdres, highres, Sparxie, 1girl, earrings, smile, looking at viewer, choker, earrings, gloves, white background, smile, cute, playful, hand on hip, simple background

全身可尝试添加full body, boots 等。

使用建议

• 服装与配饰 ： 针对服装 hard-baked问题进行了专门优化。现在官方外设是可拆卸的，官方服设主要绑定在绑定了以下标签：hat, dress, choker, gloves, boots, earrings。

• 更换服装： 实测下来我在我常用的checkpoint模型上在去掉上述官方服装提示词后，可以自由添加任意喜欢的服装元素。如果你在使用其他checkpoint模型推理时仍发现有官方服装残留，建议将上述标签加入到负面提示词中。

小记

虽然我已经对火花的特征还原程度很满意，包括眼部装饰、小礼帽、耳饰等等，但不能保证所有细节稳定，我稍后会发布细节优化的版本。

如果有任何反馈意见欢迎评论，以便我做出调整，谢谢。我会定期在社交媒体发布图片，可在对应平台取原图。

Bilibili账号：https://space.bilibili.com/12595237?spm_id_from=333.1007.0.0

小红书账号：https://www.xiaohongshu.com/user/profile/68483ecb000000001b019555