No, I’m with you on this one - this creator has made several models, but they’re all embeddings when similar models have been LoRA's. I’m also curious as to why they’re embeddings instead.

I'd counter with the question: why are concepts/characters that could be 10-12KB embeddings 217MB LoRa instead? Why re-teach a model something it's already seen and just needs words to describe in a cumbersome, slow and data-hogging way by creating an entirely new set of weights?

FFrince No. A LoRa is a collection of trained weights, a discrete network applied to the underlying checkpoint model. That network may introduce new concepts or refine existing; lighting up that new network may require the use of keywords or not.

An embedding (textual inversion) is simply a set of vectors that tell underlying models which pre-existing layers to light up.

For concepts, characters, styles or locations that are already well-known to the base model, creating a LoRa to overlap that is a waste of space; the proper/efficient method is to simply tell the model what to express. E.g., it'd be silly to train a LoRa on the concept of "1woman"; it's baked into every single SDXL-derived model already.

In this specific case, lying prone, sex, 1man and 1woman are already well-known concepts. Creating a LoRa to capture this idea, rather than an embedding, would be silly. Hence my counter-question.

FFrince And to clarify, I agree with your tactic to create an embedding rather than a lora for this concept. It's exactly right. My response was mostly to ProphetofDoom31 and Immersturm.

sendmefreestuffa8246 awesome, great info! Thanks for the reply and explanation, glad to know the differences! :D