鳴海兎亜 Narumi Toa (鸣海兔亚) | ワガママハイスペック WAGAMAMA HIGH SPEC (任性HighSpec)

https://civitai.red/models/2828908

来自任性HighSpec的 鸣海兔亚~ / Narumi Toa from Wagamama High Spec~

使用了CAME 优化器在低学习率下学习，对细节还原有一定增强 / Trained with the CAME optimizer at a low learning rate for better detail retention.

我的训练器/ Trainer：amenorira/lora-scripts-anima

Trigger Word (触发词): 只训练了校服和私服Trained on her school uniform and casual outfit only.

校服 / School Uniform：narumi toa, red eyes, long hair, school uniform, long white sleeves, hairband, bow, skirt, pleated skirt, hair bow, white hair, black pantyhose

增强校服裙子细节 / Extra School Uniform Skirt Details：white hem stripe, pink piping, dark red pleated frills, layered contrast hem

私服 / Casual Outfit：narumi toa, red eyes, long hair, hairband, bow, hair bow, white hair, thighhighs, frilled black dress, frills, black thighhighs, blush, over-kneehighs, frilled thighhighs, long sleeves, zettai ryouiki

增强私服裙子细节 / Extra Casual Skirt Details：small black bows on the waist sides, vertical white lace trim, layered skirt with white lace trim, dark grey and black diamond plaid pattern panel spliced at the lower skirt

上面提供的“增强细节”提示词并不是在训练时就加进去的。正如我在我的另一个lora（薇莉丝 Violet | 满级穿越的我为什么是牧师小姐啊 I Am a Max-level Priestess in Another World）里所说的，这些提示词是训练完成后的提示词工程所发掘的，利用anima的自然语言理解能力，能更好还原一些细节。如果你也对anima的lora训练有兴趣或者正在训练anima的lora，不妨也尝试下。

如何简单地发掘这些提示词：在训练后的测试过程中，用生成的图对比人物原图，找出你觉得还原得不好的地方，在原图上圈起来发送给任何具有图像理解能力的AI Agent（我使用的是Deepseek V4 识图模式），询问“这个东西用于ai文生图应该怎么描述”。

The “extra detail” prompts above were not included in the training captions. As mentioned in my other LoRA（薇莉丝 Violet | 满级穿越的我为什么是牧师小姐啊 I Am a Max-level Priestess in Another World）, I discovered them during post-training testing. Anima’s natural-language understanding can help recover finer outfit details. If you are interested in training Anima LoRAs, give this method a try.

A simple way to find these prompts: compare your test generations with the original artwork, mark any details that are not reproduced well, and send the marked reference to a vision-capable AI assistant. I use DeepSeek V4 in vision mode and simply ask: “How should I describe this detail in an AI image-generation prompt?”