为什么你的AI角色面部总在变化？

你写了同样的提示词两次，却得到了两个不同的人。这不是模型固执——而是模型在做它被设计来做的事。在文生图流程中，没有任何东西会把你的角色从一次生成带到下一次，除非你刻意携带。一旦你明白哪些输入会影响身份，修复方法就显而易见了。

模型不知道你的角色是谁

生成是条件采样：提示词+种子+权重输入，从海量可能的人脸分布中输出一个样本。"玛雅，28岁，深色头发，绿色眼睛"描述了成千上万张相似的脸。每次生成都会在该空间中选取一个点。任何输入稍有变化，你就会落在不同的点上——通常是不同的人。

真正导致两次生成之间人脸变化的原因

• Seed — different initial noise produces a structurally different person before your prompt has any say.
• Prompt wording — synonyms, adjective order, and added details all shift conditioning. "short bob" vs "bob haircut" is enough.
• Model or route — different checkpoints and samplers carve the latent space differently; the same words describe different people on each.
• Resolution and aspect ratio — coarse features like skull shape get re-decided when the canvas changes size.
• Refiner and upscale passes — second-stage models happily redraw jawlines and eyes while "improving" detail.
• Batch mode — generating many variants means many identities by design. A batch is an audition, not a library.

区分三种失败模式

• Identity drift — a genuinely different person. Fix with references, not prompts.
• Render drift — same person drawn differently: softer skin, wider spacing, changed age. Fix by editing the output (inpaint) instead of rerolling.
• Framing drift — right face, wrong crop or lens feel. Fix with composition language and camera presets, leave identity alone.

真正固定一张脸的方法

1. 01Produce one canonical portrait everyone copies from. Stop auditioning new versions of the face once you have it.
2. 02Freeze an identity descriptor block — a paragraph describing the face structurally — and paste it verbatim into every brief.
3. 03Create new scenes by reference-based editing from approved frames, so the pixels of the face travel with the request.
4. 04Repair drift with masked inpainting: only the broken region resamples, everything outside the mask stays pixel-identical.

知道何时漂移是预期的

极端的头部角度、强烈的运动模糊和刻意的风格转变总是近似而非克隆。请根据媒介设定容差——电影剧照应比200px缩略图更严格——并用针对性编辑修复异常值。追求每一帧的像素级身份一致性，消耗的积分比建立可信度更快。

接下来：完整的四层工作流，让角色在数十张图像中保持可识别性——参见本节中的“如何保持AI角色一致性”。