There are a number of possibilities, but it is possible that the trigger words used when training LoRA are not included in the prompts, or that they are not being recognized properly by the model.
Trigger words are not that special, and the main thing is to make the model remember the connection between the key word and the image. It is more convenient to use multiple LoRA when there are conditions. Sometimes, if the base model already has a strong memory of a word, it may not be possible to overwrite it, so some people deliberately use meaningless character strings as triggers. There are also LoRA that have been trained so that they don’t need trigger words. There are also LoRA that have been made using more specialized methods.
In addition, there is a possibility that the number of images used for training the model is simply not enough. In some cases, 10 to 100 images are enough to capture facial features, but if the concepts are abstract, such as “situation” or “style of painting”, then 100 to 1000 images, or even more samples, may be needed for successful training.
Alternatively, there may be a mode suited to this in the training script settings.
I think the default mode is to train appearance features, because that’s how most people use it. I’m not familiar with artistic matters, so you’ll have to look for a know-how site… there’s probably one somewhere.