触发词：“白丝”或者“白丝丝袜”或者“white pantyhose”.

这是一个使用z-image-base模型训练的lora，可以用于base和turbo模型的文生图。强度是1就可以了。

我在训练时使用了mask功能，所以不会影响其它脸部lora的效果，可以一起使用。

在使用时，偶尔会出现构图把头部卡出画框的情况，如果是在turbo模型下使用，描述一下人物表情即可。如果是base模型下使用，抽两次卡就行了。

我上传的图片既有turbo模型生成的，也有base模型生成的。我总是控制不好CFG/调度器/采样器，所以生成图片效果似乎不是太理想。但这应该不是这个lora的问题，因为我不用它，生成的图也那样……

总之，欢迎返图。

另外，我在lora训练方面还是新手，并不太熟练。我使用的工具是onetrainer，因为AITOOLKIT训练出来的lora似乎很难兼顾base和turbo模型。也欢迎大家留言讨论lora训练方面的话题，我期待能从各位的留言中获得新的知识和经验。

trigger word “white pantyhose”.

This is a LoRA trained on the z-image-base model. It works with both base and turbo text-to-image models. A strength of 1 is sufficient.

During training, I used masking, so it won’t affect other face-related LoRAs and can be used together with them.

When using it, the composition may occasionally crop the head out of the frame. If you’re using a turbo model, simply describe the character’s facial expression. If you’re using a base model, just reroll once or twice.

My sample images were generated with both Turbo and Base models. I’m not the best at dialing in CFG, schedulers, or samplers, so the results might not always look perfect. That said, it’s not an issue with the LoRA—I get similar results even without it.

Feel free to share your generations!

I’m still pretty new to training LoRAs, so I’m definitely not an expert yet. I used OneTrainer for this, since LoRAs trained on ZImageBase don’t seem to carry over very well to Turbo models (I’ve seen this mentioned on Reddit as well). They work perfectly fine on Base, but the effect can be pretty weak on Turbo.

Feel free to drop comments or share your experience with LoRA training—I’d love to learn more from you all.