十里平湖霜满天寸寸青丝愁华年对月形单望相护只羡鸳鸯不羡仙致童年女神，勾魂摄魄聂小倩。对于非白人女性Emb，参考我的出图提示，可避免白人特征的污染。因为绝大部分的ckpt本质上就是一个大型的emb词典集合。他们的作者在基于1.5引擎训练的CKPT，在关于woman或相关的其他单词必然是基于白人为训练模本的（1.5引擎本身也是如此，双重暴击），所以只要你的提示词里包含有类似：man、woman、naked、nipple等和人类、人体器官相关的单词，一定一定会出现白人特征，连阴毛和生殖器都是！另一方面，如果你尝试加入限定词汇asian或其他种族，又会影响角色EMb的使用，因为出来的效果都是带有种族歧视刻板印象的高浓缩向量，反而会影响其他插件模型或角色模型的使用效果。解决的方法有两种：1、使用高浓缩向量密度的emb或lora，以此来覆盖掉白人词汇的向量影响力。全看作者炼丹能力，但在怎么努力，也无法完全抹平底模引擎天生的种族缺陷。2、参考我的提示词，用双重确认的方式+高浓缩向量密度emb，这个方法可以完美解决人类单词被白人特征独占的问题。PS：我的作品都是高浓缩向量密度。双重确认不是加括号提高权重，提高权重毫无帮助。For non white female Emb, refer to my graphic prompt words to avoid contamination of white features.Because the vast majority of CKPT is essentially a large collection of EMB dictionaries. Their author is training CKPT based on the 1.5 engine, and other words related to woman or others must be based on the training template of white people (the 1.5 engine itself is also the same, double critical strike). Therefore, as long as your prompt words include words related to humans and human organs such as man, woman, naked, nipple, etc., white characteristics will definitely appear, even pubic hair and genitalia!On the other hand, if you try to add the restricted word Asian or other races, it will affect the use of character EMb, because the results are highly concentrated vectors with racist stereotypes, which will affect the use of other plug-in models or character models.There are two ways to solve it:1. Use highly condensed vector density EMB or LORA to cover the vector influence of white vocabulary. It all depends on the author's alchemy ability, but no amount of effort can completely eliminate the inherent racial flaws of the bottom mold engine.2. Referring to my prompt words, using double confirmation and high condensed vector density emb can perfectly solve the problem of human words being monopolized by white features.PS: My works are all highly condensed vector density. Double confirmation is not adding parentheses to increase weight, increasing weight is not helpful.

Files with SHA256: f82eb73cead499947fd19835c3769fdff0f04aefb677df1b1ec5249e5cee05cf