来自任性HighSpec的 鸣海兔亚~ / Narumi Toa from Wagamama High Spec~使用了<code>CAME</code> 优化器在低学习率下学习，对细节还原有一定增强 / Trained with the <code>CAME</code> optimizer at a low learning rate for better detail retention.我的训练器/ Trainer：<a rel="ugc" href="https://github.com/amenorira/lora-scripts-anima">amenorira/lora-scripts-anima</a><hr />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：<code>narumi toa, red eyes, long hair, school uniform, long white sleeves, hairband, bow, skirt, pleated skirt, hair bow, white hair, black pantyhose</code>增强校服裙子细节 / Extra School Uniform Skirt Details：<code>white hem stripe, pink piping, dark red pleated frills, layered contrast hem</code>私服 / Casual Outfit：<code>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</code>增强私服裙子细节 / Extra Casual Skirt Details：<code>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</code><hr />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（<a target="_blank" rel="ugc" href="https://civarchive.com/models/2630642/violet-or-i-am-a-max-level-priestess-in-another-world">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</a>）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（<a target="_blank" rel="ugc" href="https://civarchive.com/models/2630642/violet-or-i-am-a-max-level-priestess-in-another-world">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</a>）, 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?”

Files with SHA256: 63cb9e0fae641014e223b3379902ffc890b441089ae4d5018142110a6b7ffa2a