That's because it's meant to activate from the trigger word. Any image associated with 'Khyle' during training will be done exclusively in that style. Per CivitAI's trainer and what it does for "trigger words": Word that serves as an "activator" for your LoRA during generation.

Another thing to consider is that if a model doesn't already know what - let's say an apple is, (most checkpoints HAVE some semblance of it cooked in, but not all of them), then you need to tell the checkpoint that you're teaching it to make an apple during training.

If a model already knows what an apple is, but doesn't know it too well or learned from low quality sources, you need to tag the images as apple, to reinforce what it already knows an apple is. The reason a lot Loras work without keywords, is because their training data might have been tagged with something really common in prompts, like 1girl, a picture of, etc. Other just work simply by the way Loras alter or hook to the base checkpoint, even if they don't have prompts related to their trained tags.

So basically, <lora:filename:strength> just tells the UI "yes, load this Lora into memory and apply it to the model at this strength". If you just type "apple" without the Lora loader syntax, the UI doesn't load anything else except the checkpoint that is already loaded (and whatever other extensions load, only if also using them).

I'm sorry it doesn't work without a trigger word, but that's how I've trained this model and frankly all of my models so that it doesn't trigger anything by mistake that shouldn't be triggered inside a baked model. If you have an issue with that, that's okay.

@rypoced463 Gotcha. Doing my own test with a checkpoint that didn't have the direct "khyle" trigger baked into the checkpoint using that keyword (https://civitai.com/models/1257951/kawaij). I ran a test using the same data (seed, checkpoint, etc.) at different strengths using the tag with and without the trigger word. (All with metadata embedded so you can run your own test: https://civitai.com/posts/23596952). In other examples ran, I found that if you encountered a model that had it pre-baked into it, the quality was rather a little more low quality whereas with the LoRA only really "improved" the quality of the image produced Khyle image produced from the base model.

Another example in that post is that my training focused on the quality of eyes produced by the model - a key thing that base Khyle produced was rounded - blurred eyes where as the lora added a more "defined" clearance to it and overall enhancement to what the baked style had. I used (https://civitai.com/models/941345?modelVersionId=1053830) to get a custom Khyle style look to support it (the close-ups of the eyes).

So in terms of strength on non-native models of Khyle it gets and supplies you with the look and to other models that have baked in khyle style to it, it adds more details and gives it a more refined look than what a base model would produce.