@cliang96844 The armour guy, what a blast from the past. : )

@diwobas662368 A specific word with which the model was trained. When the prompt contains it, the model's effect is greatly amplified. Some models don't have it and simply activate with higher weight or general description. Beyond xl, bases use llms for text encoding instead of clip, so word vomit is the unfortunate standard.

How did you tag your dataset? Did you include specific words or phrases?

I noticed in the other example images the term "tenue du 18th siècle" is used, is it the same as in your dataset

I just put some janky description like Retro 18th Century Clothing and it seemed to render the outfit from the samples? I am using CyberRealistic V10.

@yajukun using consistent keywords in your training data allows more flexible activation of lora concepts. The LoRA from my experience will always have some influence on the output, but if you tagged the data with a keyword and that keyword is not used, the concept is not actvated.

I see for your LoRA, the outfit is always activated even when other clothing are prompted (or nude). In the case of this lora it's not an issue, but with a key token and auxiliary tokens (like "ruff", "high collar", etc.), you can better control parts of the outfit.

Also the LoRA maybe over trained since openpose cannot control the final pose of the person very well.

or maybe you should tag the person's pose in your dataset, because it is more likely the LoRA considers the arms down pose as a part of the clothing concept instead of a character/pose concept.

Basically just consider tagging manually: https://civitai.com/articles/29347/the-ultimate-lora-tagging-guide-from-dirty-image-to-perfect-model

It still holds for SDXL models at least