Stable Diffusion写汉服模特图的提示词
来源:互联网
时间:2026-06-28 13:19:05
要用Stable Diffusion生成一张细节精准、风格统一、不崩形不穿模的汉服模特图,光写"hanfu"或"Chinese dress"是远远不够的——模型大概率会随机套用明清戏服、旅游拍照装,甚至把唐风襦裙和清宫旗装混搭起来。真正有效的做法,是建立起一套分层明确的提示词结构,从人物基础属性到服装形制、材质工艺、再到环境氛围,逐层锁定,每一步都不能含糊。

先给出核心思路:四层锚定法。这不是什么玄学,而是经过反复试错总结出的实操框架。
核心提示词结构:四层锚定法
第一步:锁定人物基础属性
1girl, 22 years old, slender but curvy, front-facing, full-body shot, centered composition, studio lighting。别小看"full-body shot"这个词,漏掉它模型会默认只画上半身,后续硬加裙摆只会强行拉伸变形。
第二步:固化汉服类型与结构
【齐胸襦裙,交领右衽,宽袖垂至膝下,腰间系双层素色绦带,下裙褶皱自然垂坠】
第三步:注入材质与工艺细节
brocade with raised gold thread,否则模型默认给你平涂一片金色块。
第四步:绑定环境与氛围
background: misty Jiangnan garden, willow branches, soft bokeh, pale ink wash tone。把背景信息塞进主提示词里,会稀释服装的权重,导致衣纹模糊、配色失真。
避坑关键词替换表
把模糊的词换成模型能真正渲染的实体描述,效果立竿见影。
比如"elegant"——直接删掉,改用standing gracefully with one hand holding sleeve;"traditional"——删掉,改用Ming dynasty style aoqun, stand-up collar, side slits on skirt;"flowing"——删掉,改用long sleeves billowing slightly in breeze, fabric physics visible。
另一个关键点:用LORA触发词强化形制可信度。加载时,必须在正向提示词中加入触发词hanfu_v30,否则LORA根本不会生效。同理,用凤冠霞帔LORA时,
【QIEMANCN必须出现在提示词最前端】
实战组合示例(可直接复制)
① 清雅宋制褙子:
masterpiece, best quality, 1girl, 20 years old, fair skin, black long hair tied with white ribbon, wearing light blue Song dynasty beizi over white ruqun, narrow sleeves, straight collar, subtle cloud-pattern embroidery on hem, standing beside ink-wash bamboo grove, soft morning light, shallow depth of field
② 明制大袖衫+马面裙:
ultra-detailed photography, 1girl, 25 years old, poised posture, Ming dynasty da xiu shan with wide sleeves and embroidered phoenix collar, red horse-face skirt with gold-threaded lotus border, waistband with jade pendant, studio backdrop of vermilion wall and lattice window, cinematic rim light
③ 汉元素改良日常款:
realistic portrait, 1girl, 23 years old, modern hanfu-inspired outfit: na vy blue cross-collar top with asymmetric knot, high-waisted grey pleated skirt, barefoot on stone path, cherry blossom petals floating mid-air, natural daylight, Fujifilm XT4 color profile