embedding only giving subject and not the rest of the prompt

[Andyholm]

So I've been messing around with the textual inversion embedding feature, but I can't seem to generate more than just the embedding subject.

Example with "me" as the embedding: me eating a hamburger

This would only return me as the subject of the photo, but nothing else about the given prompt. Am I doing anything wrong?

[garlan999]

are you putting the me in like this? eating a hamburger?

The editor removed the more than and less than symbols. The 'me' should be within less than and more than signs But I can't type it.

[Andyholm]

Hmm, not sure what you mean, sorry.
The embedding is "andreas", so if I'd want myself to eat a burger I'd write: andreas eating a hamburger
Is that not the correct way of doing it?

Same result when the prompt is: <andreas> eating a hamburger
vs: andreas eating a hamburger

[telemole]

When you run a generation - it will tell you which inversions are referenced after it's done - this is how i check to make sure it was using the expected embedding

[Andyholm]

It's using the embedding, the issue was that it's not doing the rest of the prompt at all. It might be because i trained them using the style template rather than the subject one. So I'll report back once it has trained on that template instead. :)

[Andyholm]

I might've been training this wrong, so I'll report back once I've tested some more.

[garlan999]

The only thing I've gathered so far is that a thing should be trained as an object (including a face), while a style is, well, a style . Or a concept could be a style. I'm not even convinced of that much. I'm looking at this https://huggingface.co/sd-concepts-library to see if I can get a better understanding.

[Andyholm]

Yeah, I trained it as a style by mistake at first😅
Changed it to subject now though so it has been getting better.
I think the second reason is that I've been using a vector count of 16.
I'll experiment more with lower values to see if I find a decent middle ground.

[CodeExplode]

The higher your embedding vector count, the more it will overwhelm the prompt. It's like using more words to describe your concept in the prompt.

Also, according to the author of the textual inversion paper, the more your final embedding vectors lies 'outside' of the range of values which the model was trained on, the less editable it will be.

[Andyholm]

Ahh makes sense. Thanks for the insight!
At first, it was trained as a style by mistake, so I changed it to subject.
The vector count has been at 16 both times, so that makes sense as to why it's harder to customize it.
Any suggestions for vector amounts?

[CodeExplode]

I'd try 2-6 vectors for a person, which from my brief experiences was enough (with different TI code though)

[Andyholm]

Alright, I'll see what 6 does and I'll go from there. I appreciate the help!