Durandal-Palatinus Equinox-Honkai 3rd

v3

中文

全新的训练数据，第一次尝试把两种服装融合进一个lora中。训练时保留了一部分服装标签，以获得更稳定的输出并方便换装，结果令人惊喜。推荐权重0.8，如需改变外观，只需保留daie white/daie black及头发相关提示词，并将权重降至约0.6即可。若换装时发色也发生变化，添加"blonde hair"即可。对于黑色外观，加入armored boots有很大概率生成全身像，即使添加cowboy shot也无法获得相应图像，需移除armored boots。同时可能生成其他穿铠甲的人物，解决方法是避免让她孤立地站立（standing），可添加背景、姿势、动作等其他描述。训练素材为纯黑背景，通常需要额外添加背景描述词。该模型在真实人物模型上也有一定表现效果，若懂得分层使用lora，可获得更佳图像。

English

Brand new training data, the first attempt to integrate two clothing styles into one lora. A portion of clothing tags were preserved during training to ensure more stable outputs and easier outfit changes. The outcome is truly surprising. Recommended weight: 0.8. To alter appearance, just retain "daie white"/"daie black" and hair-related prompts, and reduce the weight to around 0.6. If hair color also changes during outfit swaps, simply add "blonde hair". For the black appearance, including "armored boots" greatly increases the chance of generating a full-body image—adding "cowboy shot" won’t help, so remove "armored boots" instead. There’s also a chance other armored figures may appear; to address this, avoid having her stand stiffly (don’t use "standing"), and enrich the description with background, pose, or motion cues. The training data uses a pure black background, so background details are generally needed. The model also performs reasonably well on real human models. If you know how to layer LoRAs, you’ll achieve even better results.

tusi.art 是一个免费的在线图像生成网站，受限于电脑配置的建议尝试一下~

V2

修复了到处都是的头发问题。

fix the meaningless hair.

ps: 试了下好像真人模型也是可以的~不保证效果，因为没有真人图片

After trying, I found that it may also be applicable to real models, but I cannot guarantee the effect.

新人第一次炼丹，出现错误请多多包涵~

This is my first lora, may have many errors. Forgive me.

不允许商业用途

NO COMMERCIAL USAGE