This is a character LoRA for Neko Bea, intended primarily for character appearance, outfit, accessory, and general concept reproduction.

The training dataset was carefully collected and curated with assistance from ChatGPT image generation, including both image generation and image restoration/editing. Considerable effort was made to improve character consistency, clothing patterns, accessories, poses, expressions, and overall illustration quality. The dataset was refined to the best of my ability for this training run.

The base model used for training is a custom merge of NoobAI (Epsilon Prediction) and an Illustrious XL model. I have not performed sufficient controlled testing to determine exactly how much each base model contributes to the final result, so compatibility with other NoobAI/Illustrious-based models may vary. Compatibility with some style LoRAs may also be limited or unpredictable.

The selected LoRA was trained for approximately 3,000 steps. It may exhibit some degree of overfitting, particularly in certain character details and compositions, but this was the version I considered the best overall balance during testing.

Concept Tags

The training dataset uses the following tags to distinguish different character concepts. These concepts can generally be combined freely in prompts:

cat costume — Neko Bea's default cat costume. You can freely replace Neko Bea's default outfit with other clothing in your prompt, and the LoRA should generally handle this reasonably well.

neck bell — The bell on Neko Bea's chest. When Neko Bea is wearing a non-default outfit, you can add neck bell to the positive prompt to recall the bell.

gloves — Neko Bea's gloves. If you want the palms/hands to be exposed, adding gloves to the negative prompt may help.

tail — Neko Bea's tail.

sign — The golden sign held by Neko Bea.

pal — The small floating cat companion next to Neko Bea, corresponding to the bee companion associated with Bea's default skin. This concept is currently unreliable and should be considered non-functional.

backpack — Neko Bea's backpack. In most cases, the LoRA handles its perspective correctly, including side and rear views. However, perspective errors may still occur occasionally.

Some of these concepts may still have residual associations with the default outfit, especially neck bell and certain accessories. Prompt control is therefore not guaranteed in every combination.

Known Limitations

• The pal concept is currently not reliably controllable.

• Fine control over some clothing patterns and small character details may be inconsistent.

• The neck bell and some accessories may occasionally remain associated with the default outfit.

• Some prompts may be interpreted toward concepts that appeared frequently in the training dataset.

• Strict control over individual color patterns may not always be possible.

• Compatibility with some style LoRAs may be limited or unpredictable.

• The LoRA is not intended for NSFW generation and does not support NSFW use.

I do not currently plan to continue updating or optimizing the dataset to address these limitations. This project was primarily a learning and experimentation project, rather than an attempt to produce a perfectly optimized commercial-quality LoRA.

That said, I believe the current version is capable of handling most ordinary Neko Bea generation scenarios reasonably well. If you don't mind the limitations described above, feel free to download and use it.

Disclaimer

Neko Bea and Bea are characters from Brawl Stars, developed and published by Supercell. This is an unofficial fan-made work and is not affiliated with, endorsed by, or sponsored by Supercell. All related character names, designs, and intellectual property belong to their respective rights holders.

这是一个针对 Neko Bea（招财猫贝亚） 制作的角色 LoRA，主要用于还原该角色的外观、服装、配件以及整体概念设计。

训练数据集是在 ChatGPT 图像生成与图像修复/编辑工具的协助下精心收集与筛选的。为了提升角色一致性、服装图案、配件细节、姿势、表情以及整体插画质量，我付出了相当大的努力。在本次训练中，数据集已被精炼至我目前能力所能达到的最佳状态。

训练所使用的底模（Base Model）是一个将 NoobAI (Epsilon Prediction) 与 Illustrious XL 进行融合的自定义 Merge 模型。由于我尚未进行足够严谨的对照测试，无法确切断定两个底模各自对最终生成效果贡献了多少，因此该 LoRA 与其他基于 NoobAI 或 Illustrious 衍生底模的兼容性可能会有所差异，与部分画风 LoRA 的兼容性也可能受限或不可预测。

最终选出的这个 LoRA 版本大约训练了 3,000 步（steps）。它在某些角色细节和构图上可能会表现出一定程度的过拟合（Overfitting），但在测试过程中，我认为这个版本在整体表现上取得了最佳平衡。

概念标签（Concept Tags）

训练集使用以下 Tag 来区分不同的角色概念。这些概念通常可以在 Prompt 中自由组合：

cat costume —— Neko Bea 的默认猫咪服装。你可以在 Prompt 中自由为 Neko Bea 更换其他服装，LoRA 通常能够比较好地适应。

neck bell —— Neko Bea 胸前的铃铛。当 Neko Bea 穿着非默认服装时，可以将 neck bell 加入正面提示词，以召回铃铛。

gloves —— Neko Bea 的手套。如果希望露出手掌/双手，可以尝试将 gloves 加入负面提示词。

tail —— Neko Bea 的尾巴。

sign —— Neko Bea 手持的金色金牌。

pal —— 悬浮在 Neko Bea 身旁的小猫伙伴，对应 Bea 默认皮肤中的蜜蜂伙伴。目前这个概念的控制效果不可靠，应视为无效 Tag。

backpack —— Neko Bea 的背包。大多数情况下，LoRA 能够正确处理背包的透视关系，包括侧面和背面视角。不过在部分情况下仍然可能出现透视错误。

注：部分概念可能与默认服装存在残留的特征粘连（尤其是 neck bell 和某些特定配件），因此无法保证在每种组合下都能实现 100% 的精准控制。

已知局限性（Known Limitations）

• pal（伙伴）概念目前无法稳定控制。

• 对部分服装图案和小细节的精细控制可能会存在不一致。

• neck bell（胸前铃铛）及部分配件偶尔仍会与默认服装产生绑定。

• 部分提示词可能会偏向训练集中高频出现的概念/构图。

• 可能无法始终做到对单独颜色图案的严格控制。

• 与部分画风（Style）LoRA 的兼容性可能受限或不可预测。

• 本模型并非为了生成 NSFW 内容而设计，且不支持任何 NSFW 用途。

• 我目前暂无计划继续更新或优化数据集来解决上述局限。本项目主要是一个学习与探索性质的尝试，并非为了制作完美无瑕的商业级 LoRA。

尽管如此，我相信当前版本已经能够相当不错地应对绝大多数日常的 Neko Bea 图像生成需求。如果您不介意上述局限性，欢迎随时下载体验！

免责声明（Disclaimer）

Neko Bea 与 Bea 是由 Supercell 开发并运营的游戏《荒野乱斗》（Brawl Stars）中的角色。本模型为官方未经授权的同人二创作品，与 Supercell 无任何关联、赞助或背书关系。所有相关角色名称、设计及知识产权均归其对应的版权所有者所有。