Using words that are never pre-populated in the model helps to avoid the side effects of ordinary words and can also help to debug their effects without the interference of the meaning of the words themselves.

My guess.

I've found that using a trigger word that is the same as or contains 1-or-more tokens that SD already knows can lead to strange results. With TI embeddings, changing the filename changes the trigger... but I don't think that works with LORAs.

@chromesun TI embeddings are in pure word form, which is confusing because it is the exact same form as the prompt. lora filenames are included in <>, so it doesn't matter, but some lora need specific prompts to be effective, and these prompts can be confusing if they are not made into very specific non-words.

Even I have tested (and readed a lot) I found that sometimes unique tokens work and sometimes not. For example, if you use "atr3" as token in some tranings works beter than others and even using unique tokens in some trainings affects in other way... I suspect tokenization + internal system of how ai is understanding what it's seeing in it's own words.

rename the lora filename with the trigger words

Generally, to avoid the influence of the words "hidden" in the trigger word.

This is a LoRa, you shouldn't need a trigger word at all. Either you want the style or you don't, and it should apply to the prompt regardless. Requiring a trigger word defeats the purpose of a LoRa - use an embedding instead.

EDIT: Now that said, this user's LoRas don't appear to need trigger words. I found this user from TentacleHorrorAI and didn't even know there was a trigger word and still got great results (I did use the word 'tentacles' though, because that 's what I was going for)

@ShivaFang You can use trigger words on LORA's to activate certain characteristics of the training via txt captions, it depends a lot of how was it trained and what you wanted. I suppose each do things on their way but I use often trigger words in LORA makings. And sometimes not.

you can click the little button "read meta data" in the lora selection on auto/others and it will show you all the words used in training.

because your prompt is bad

(red_human_heart_organ), in a glass jar <lora:kVoidEnergy-000001:0.8> v0id3nergy, / violet light studio lighting, fantasy, glowing

in the neg put ((yellow light)), purple object