Best way to add custom html elements in extensions? (non-gradio)

[CodeExplode]

I'm working on a manual embedding editor, and want to have tagged positions on the weight sliders for user-provided words (tokens).

Since you can just combine embeddings to get a new valid embedding (e.g. a skunk-puppy hybrid), it seems plausible that manually moving weights between marked embeddings might help find and fine-tune exactly what you're looking for rather than rely on textual inversion, though it's possible that relative relationships between the weights are just as important and moving one weight at a time won't work, but I'd still like to try. Having a bunch of animals tagged on each weight might give an idea of valid ranges if you're trying to create say a pokemon which will still work like an animal (or create a starting point for textual inversion, since there are multiple valid embeddings for a given output but some will have better editability than others).

Gradio slider are a bit bulky for 768 weights, and afaik there's no way to add markers on them for the user-provided words. I'm happy to add my own sliders, but don't know the ideal way to do it in the extensions framework which wouldn't be considered an ugly hack.

[CodeExplode]

An example of the indicators I'd like to build are like this (may not be the most colour-blind friendly choices):

Currently if you were to have a prompt with 'red shirt' and 'blue castle' and 'orange hair', the colours would tend to get mixed up. But if you used an embedding which already has a colour implied within it, like apple or tree or ocean, it doesn't tend to bleed into the colour of everything else. So maybe a better way of getting a red shirt would be to make a tweaked version of the shirt embedding to always produce red shirts, and maybe the way to get there would be to look at the bounds of other red objects (and ideally a red piece of clothing, but I can't think of any off the top of my head), and create a single vector embedding for 'red shirt' to then use in a prompt.

So 'a man wearing a red shirt standing in front of a blue castle' could perhaps be done with 'a man wearing a redshirt standing in front of a bluecastle' where those new embeddings were found with some guided tweaking, and there's no standalone colour descriptors to bleed into the rest of the prompt.