santa hat
deerdeer nosedeer glow
Sign In

Random Schizo Ideas of a clueless user - Embedding Merge something something


Filler Title To Catch Your Attention

Hear me out (untested)

Clip Interrogator space on Huggingface can give you a very precise description of any picture you feed it, and that description is pretty much how any SD1.5 would understand the picture in itself.

Let's say you have X pictures of a character you want to make a textual inversion with.

Feed all those pictures to the Clip Interrogator and get both the positive and negative description/prompt of them.

When you have all the descriptions/prompts, download the Auto1111 extension "Embedding Merge" by klimaleksus.

Use the extension as to build up discrete embeddings, one for each positive description/prompt and one for each negative.

Use the extension to merge every positive embedding in one single positive embedding and every negative embedding in one single negative embedding (using the plus, +, expression syntax)

After the first merge, merge again the two resulting embeddings but with a minus, -, expression syntax, in order to 'remove' the negative embedding values from the positive embedding, having a resulting final embedding.

Test the final embedding and see if it has resemblance with the subject you wanted, if it REALLY works then try to resize the final embedding, since it will probably be very large.

"Yo, why should I work this hard for a crappy textual inversion?"

* cries in 4gb Vram *