Mix Embedding and Model

[msoucy]

Hey, I had a question about Textual Inversion embeddings and models.

From a quick test, I'm able to use an embedding trained with one model, for generation with a different model. It could just be the model I was using (classic Disney animation style), but I appeared tinted red.

I thought I saw somewhere that you're not supposed to mix embeddings with a different model, though? Is that still the case? How does this work with posts I've seen saying "the best result uses Dreambooth and embeddings together"?

[ghost]

Embeddings are trained using specific model, and work best with only them. They are possible to use with any other, but results will be worse, because it's not really compatible with them. Same goes for hypernetworks. Dreambooth is fine-tuning model, and as i understand you can just use it to train embeddings, and it's not changing model completely, so it probably works fine with embeddings that were trained on non-dreambooth model.

[msoucy]

Thank you, that sounds like what I had heard. It also kinda makes sense as to why an embedding sort-of works with a different model, if that model is derived from the same model that the embedding is.