用liblibai做人物设定需分三步：先建“角色身份证”锁定面部锚点，再通过controlnet扩展全身并严控脸部一致性，最后用毫米级动作、接触关系和中断点定格行为逻辑，确保角色可复现、可延展、可验证。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 多模态理解力帮你轻松跨越从0到1的创作门槛☜☜☜
用LiblibAI做人物设定时，必须把“同一个人”从模糊印象固化为可复现、可延展、可验证的视觉资产，不能只靠一张图蒙混过关——漏掉发根蓬松度或袖口褶皱逻辑，第二张图里角色就可能长出三只手、穿错季节衣服、甚至瞳孔颜色突变。
第一步：先做“角色身份证”，锁定面部锚点
打开LiblibAI，新建提示词框，第一句必须写：“front-facing headshot, upper body, pure light gray background, even studio lighting, sharp focus on eyes and facial structure”。这句是强制AI放弃所有环境联想，只专注人脸本身。
紧接着填入五项不可省略的物理描述：①【性别与年龄段】写“female/26-29”或“male/33-37”，不能写“青年男性”；②【脸型与骨骼特征】写“oval face, defined jawline but soft submental area, slight nasolabial fold when neutral”；③【眼睛细节】写“brown irises with radial striations, upper eyelid crease slightly deeper on left, natural tear duct highlight”；④【发根状态】写“slight crown lift at hairline, roots slightly lifted by static, no oil sheen”——【发根状态是关键锚点】，模型对它极其敏感，漏掉就会生成洗吹后紧贴头皮的假发感；⑤【皮肤真实缺陷】写“two faint freckles on right cheekbone, one shallow acne scar near left temple, pores visible only on nose and T-zone”。
负向提示词必须含：“(deformed), (disfigured), (bad anatomy), (extra limbs), (text, logo, watermark), (asymmetrical eyes), (crooked nose), (blurry face)”——其中“(blurry face)”不可删，否则AI会默认柔焦处理，导致后续全身图无法对齐五官比例。
第二步：扩展全身，但脸必须纹丝不动
复制上一步生成的“身份证图”作为参考图上传到Img2Img模块，启用ControlNet中“OpenPose+Face”双控制。
正向提示词开头加：“full-body portrait, same face as reference image, front view, standing pose, balanced weight on both feet, arms relaxed at sides”。这句话不是修饰，而是指令——AI若没看到“same face as reference image”，就会在全身构图中自动优化脸型，导致漂移。
接着补全四层结构：①身高体型：“168cm, hourglass figure, shoulder width 38cm, waist-to-hip ratio 0.72”；②服饰穿法：“oversized oatmeal knit sweater→left sleeve pushed to mid-forearm→right sleeve hanging loose→hem falling 5cm below iliac crest”；③动态细节：“right hand holding ceramic mug, thumb resting on handle ridge, index finger lightly pressing mug rim, steam rising in thin curl”；④材质与附着：“sweater wool fibers slightly pillaged at elbows, mug surface shows matte glaze with one hairline crack near base”。
负向提示词追加：“(different face), (face altered), (reconstructed face), (plastic skin), (unnatural proportions)”——这些词专治AI在全身图中擅自“优化”面部轮廓的坏习惯。
第三步：定格角色行为逻辑，让动作有因果
方法一：用毫米级动作绑定呼吸节奏。写“woman squatting to pick up dropped notebook, left knee flexed 105°, right foot flat, diaphragm expanded causing sweater hem to rise 1.2cm above waistband, notebook corner pressing into left palm creating 0.3mm indentation”。没有角度和形变量，AI只会套用标准蹲姿模板，脸又开始漂移。
方法二：用接触关系锁死道具逻辑。写“holding stainless steel kettle with right hand, thumb on spout lever, index finger curled under handle curve, kettle base resting on left forearm, slight pressure dent visible on forearm skin”。这里“pressure dent”是关键，它迫使AI计算真实重量传导路径，连手臂肌肉走向都会被带出来。
方法三：加入不可逆中断点强化真实感。写“notebook page fluttering mid-air, top corner caught by draft, trailing motion blur toward bottom-right, one paperclip detached and rotating clockwise”。这句不单是画面描述，更是告诉AI：“此刻正在发生”，从而抑制静帧式摆拍感。
