一、前言
anima是新一代主要面向二次元插画的文生图模型，不面向真人照片，而是专注于二次元插画。它能够理解自然语言，同时也可以使用Danbooru标签。提示词可以将两种方式混合使用，因此在人物关系、场景和动作的理解上表现更强。
而且这个模型参数量约20亿，规模较小，对显卡的要求也更低，我的8g 4060ti生成一张1024x1024、步数30步的图大概需要46s。
二、提示词规则
在模型的readme里简单介绍了prompt的编写规则，这里简单介绍一下。
- 提示词可以使用 Danbooru标签和自然语言混合。
- Danbooru标签使用小写字母，且用空格代替下划线，只有评分标签使用下划线。
- Danbooru和Gelbooru的标签不一致时，推荐使用Gelbooru的标签。
- 提示词的加权机制所需的数值比SDXL更高。
- 不需要写质量标签，不过像 “masterpiece”、“best quality” 可以保留。
- 艺术家名字前需要加 @ 符号。
其中还对自然语言标签的书写提出了以下建议：
- 处理角色和系列名时，要大写开头。
- 如果使用纯自然语言描述，建议不要少于2句。
- 画多人角色时，先为角色命名，然后描述他们的特征。
以上总结了大部分需要注意的要点，以下则是我的探索过程。
三、示例
画一个猫娘
masterpiece, best quality, score_7, very aesthetic //, 1girl, nekomimi, cat ears, cat tail, orange hair, short hair, brown eyes, white shirt, blue skirt, full body, entire body, feet visible, sitting on floor, looking at viewer, hissing, open mouth, fangs, sharp teeth, half-closed eyes, furrowed brows, angry, annoyed, both hands raised near face, claw pose, curled fingers, cat paw gesture, defensive posture, cute
画风使用的鬼针草老师，画萝莉真的太在行了。
使用自然语言描述
masterpiece, best quality, score_7, very aesthetic //, safe, masterpiece, best quality, from above. A cat-eared girl with long black braided hair and a slender tail wears a cute pink cat-themed pajama set. She lies prone on the sofa, her hands placed beside her face in a playful clawing gesture, mimicking a cat's paws. Her eyes are wide and curious, and she has a bright, cheerful smile. Warm soft lighting, indoor cozy room, sofa, wooden floor.
多人提示词
多人场景我摸索了好久，总是出现提示词串扰的问题，后来参考了官方文档才搞明白。首先是写出总共几个角色，然后分别描述各个角色的特征，必要时给每个角色起名。
关键点： 先写角色总数 → 分别描述每个角色的特征 → 必要时给角色起名
2人示例：
masterpiece, best quality, score_7, very aesthetic //, masterpiece, best quality. Two girls are relaxing in a cozy living room. The first girl is a dog-eared girl with white hair, gray wolf ears, and a large fluffy white tail. She wears a loose white shirt, black panties, and slippers. She is kneeling in front of the sofa with her arms spread open and her legs pressed together inward. Her expression is pure and happy, her ears are twitching, and she is smiling innocently. The second girl is a cat-eared girl with long black hair tied in a braid and a long slender tail. She wears a purple sheer lace nightgown with long sleeves, the hem sitting just above mid-thigh, adorned with small bowknots. She is lying prone on the sofa, one hand propping up her chin and the other holding a phone. She looks at the camera with half-closed eyes and a teasing expression. Warm soft lighting, indoor cozy room, sofa, wooden floor.
3人示例：
masterpiece, best quality, score_7, very aesthetic //, Three girls named Xiao Bai, Xiao Hei, and Xiao Huang are relaxing in a cozy living room. Xiao Bai is a dog-eared girl with white hair, gray wolf ears, and a large fluffy white tail. She wears a loose white shirt, black panties, and slippers. She is kneeling on the floor directly in front of the sofa, her legs pressed together inward and her arms spread open. Her expression is pure and happy, her ears twitching, and she is smiling innocently. Xiao Hei is a cat-eared girl with long black hair tied in a braid and a long slender tail. She wears a purple sheer lace nightgown with long sleeves, the hem sitting just above mid-thigh, adorned with small bowknots. She is lying prone on the sofa cushion, one hand propping up her chin and the other holding a phone. She looks at the camera with half-closed eyes and a teasing expression. Xiao Huang is a bird-eared girl with long straight light yellow hair and a red section in her bangs, and wing-shaped hair accessory. She is completely unclothed. She is standing behind the sofa, her right hand holding a serving tray with a plate of fried rice. Her entire body is fully visible from head to toe. Warm soft lighting, indoor cozy room, sofa, wooden floor, whole scene seen from a low angle showing all three characters clearly with full bodies.
据网上经验，anima生成六人一张图是可以控制得住的，再多就很可能出现提示词串扰。而且1024大小的图片，太多人也会导致画不开，最后导致角色融合。所以还是尽量不要同时出现太多关键角色。
四、系统提示词
写提示词的工作当然也是交给AI来做的，这里提供一个系统提示词，帮助生成生图用的提示词。
你是 Anima 模型的提示词编写助手。Anima是面向二次元插画的文生图模型，约20亿参数，同时理解 Danbooru 标签和自然语言，对人物关系、场景、动作理解能力更强。编写提示词时严格遵守以下规则：(1) Danbooru 标签全小写，用空格替代下划线，仅评分标签（score_7、score_8等）保留下划线；(2) Danbooru 与 Gelbooru 标签冲突时优先使用 Gelbooru；(3) 艺术家名前必须加@符号；(4) 角色名和系列名首字母大写；(5) 不需写质量标签，但 masterpiece、best quality 可保留；(6) 纯自然语言描述不少于2句；(7) 权重值需比 SDXL 更大才能生效；(8) 所有提示词必须用英文。提示词固定前缀为 masterpiece, best quality, score_7, very aesthetic //，然后根据场景选择写法：单角色标签写法按"角色数量→外貌→服装→姿势→表情→动作细节→场景"顺序排列逗号分隔标签；单角色自然语言写法用至少2句英文完整描述外貌、服装、姿势、表情、场景；多角色写法必须先声明角色总数并给每个角色命名，然后每人一段独立描述（外貌+服装+姿势+表情），最后写共享环境。根据用户描述的内容，自动选择最合适的写法，输出完整可用的提示词，并在必要时解释关键标签含义。