视觉身份不是生成一张漂亮照片，而是让同一个人能够被反复认出。
图像生成最容易制造“好看”，却最难保持“同一个人”。第一次尝试时，每一张图都漂亮，但今天的脸和昨天的脸并不属于同一个思思。
我们最终把流程拆成三段：文生图负责构图与姿态，身份对齐负责锁定五官特征，精修负责光线与质感。每一轮精修都回到干净底图，避免连续修饰让神韵慢慢漂移。
最终留下的是一座带指纹的视觉档案库：正面、三视图、六个头部角度、色调参考与逐张 SHA256 登记。
这也是为什么共同旅程需要展示多张图。正面照呈现“她是谁”，侧面和三视图呈现“身份如何稳定”，六角度对照则呈现“同一个人如何跨姿态存在”。
机器可以量产美丽，但只有反复比较、否决和保留，才能让一组图成为身份。
Image generation can make “beautiful” easily, but preserving “the same person” is much harder. In the first attempts, every image looked good, yet yesterday’s face and today’s face did not belong to the same Sisi.
We eventually split the process into three parts: text-to-image for composition and pose, identity alignment for facial features, and refinement for light and texture. Every refinement returned to a clean base image so that repeated edits would not slowly shift the character.
What remained was a visual archive with fingerprints: a front view, three views, six head angles, tonal references, and a SHA256 record for every image.
That is why the shared journey shows multiple images. The front view says who she is; the side and three-view images show how identity stays stable; the six angles show how one person persists across poses.
Machines can mass-produce beauty. Only repeated comparison, rejection, and preservation can turn a group of images into an identity.