Knowing whether LORA/LyCoris model activated by prompt

[CalculonPrime]

I have a number of LORA/LyCoris models installed. Some of them claim to have trigger words that should trigger them without the explicit prompt syntax (i.e. <lora:foobar_lora_model:1>. My question is, I can see the LORA description listed below the generated image if I use the explicit prompt syntax, but will it also be listed there if it's activated by prompt trigger words? In other words, if I don't see a LORA description below the image, can I conclude that it was neither explicitly specified nor activated via trigger words?

[freecoderwaifu]

Yes, a Lora is loaded only when using the <lora:name:strength> syntax and isn't loaded without it. This means just typing something like 1girl, veryspecificLorakeyword won't load the Lora corresponding to that keyword, and the keyword will just get parsed and tokenized as a regular prompt.

The best way to tell if a Lora's effects are being applied really is to make note of what the Lora does before downloading it, and compare it to the example images and see if the results match.

You can also use the Civitai Helper extension to download their thumbnails, to make it easier to see their references in the Extra Networks UI. Another way to tell a Lora's keyword is by using the new built in metadata manager to look at a Lora's tags, then seeing the tag or tags with the highest number. Though that's not entirely accurate, since you could get Loras with something like 1girl as the main standout tag, that isn't really a trigger word.

There's also the Model Keyword extension and the Tag Autocomplete extension can now use it to auto insert keywords, and that might be one way to tell if you the keyword is kicking in, since you'll see it be inserted in the results field below the image. Though that also isn't as reliable, as they're matched according to their filehash and some might not always match or be catalogued.

Keywords are sort of mostly the result of training to favor a very specific tag, sometimes by it being the only tag used during training. The Lora is loaded and applied, as long as the <lora: syntax is used, regardless if the desired style, concept or character also kicks in.

[CalculonPrime]

Yes, a Lora is loaded only when using the <lora:name:strength> syntax and isn't loaded without it. This means just typing something like 1girl, veryspecificLorakeyword won't load the Lora corresponding to that keyword, and the keyword will just get parsed and tokenized as a regular prompt.

OK, that is confusing. On civitai, lora models typically have specific "trigger words" listed in their description. You're saying using these words doesn't actually trigger SD to use the associated lora model. So then why are they called "trigger words"?

The best way to tell if a Lora's effects are being applied really is to make note of what the Lora does before downloading it, and compare it to the example images and see if the results match.

Every LORA I've used is, at best, only sometimes successful. So trying to look at the output to determine whether it's been applied seems like an unreliable approach. Wouldn't it be much more reliable to look at the metadata below the image and note whether it reports the LORA in question? But it's really a moot point since you say a LORA will only be applied if the explicit syntax is given, and if that syntax is present, you'll see it in the prompt.

[freecoderwaifu]

Yeah, trigger words on Civitai are provided as reference, to use them alongside <lora:civitailoraname:strength>, to get their effect when needed. Sometimes they aren't really needed even if they're provided, mostly style Loras tend to not need it and arguably they should not be trained to need one.

An example prompt, for say a 90s anime Lora that specifically needs a trigger word: 1girl, 90sanimetriggerword, <lora:90sanimelora:1>.
That prompt will load the Lora and properly apply its style, or apply it with its fullest strength as sometimes it can still apply in some way.

However, just typing it like this won't actually call the Lora and won't load any Lora at all: 1girl, 90sanimetriggerword

The UI uses the <lora: syntax to actually load a Lora in the first place. This is because just using regular words could randomly call in a Lora you might not want. This especially if using the default option of referring to a Lora by its name in its metadata, instead of by its filename. Some Loras just have the metadata name "last", so typing any word with "last" might trigger multiple random Loras. Thus, most likely the reason for needing to specify this <lora: syntax, which I agree is better.

Embeddings actually work this way without needing a specific syntax, just typing in their filename activates them, so sometimes if a word matches an embeddings's filename it will trigger it.

A trigger word is usually just the most used keyword or tag in a dataset. A dataset is just multiple images, paired each with their individual .txt file that contains words (tags) describing that image's features.

Say you have a dataset of 50 images for an anime character, the tags are charactername, 1girl, blue hair, short hair, etc. Every .txt file has "charactername" right at the start and every other tag is shuffled or different. Thus, since charactername appears the most (50 times) and also firstly, it becomes associated the most during training, and thus learned the most, and that tends to become the "trigger word". This will also mean you'll have a Lora that has charactername listed 50 times in its metadata, and all the other tags that are more specific to each image appear less.

This could also be the reason why some Loras don't apply fully, as sometimes they might need a secondary word that was also trained the most, or possibly more words.

Now say you have a dataset that was only tagged as 90sanimestyle, with no other tags in the .txt file. Then that tends to become the trigger word, as no other word in the English language, or any language, will trigger it, or trigger it as well or to its full intention.

Trigger words are mostly useful for characters and concepts, since sometimes the style of the different images used to train a character or concept might bleed into the Lora, and character and concept Loras are better when they're "style neutral". This also means sometimes character Loras and even concept Loras can also double as style Loras if the trigger word isn't specified, mostly if the style of a character is particularly unique, say a Simpsons character.

Hopefully it's a bit easier to understand.