santa hat
deerdeer nosedeer glow
Sign In

LoRA's or Embeddings? Which do you see as the superior technique for Concepts and Characters

Just wanted to get the community's opinion on what they prefer using and/or which is the superior method in their experiences. Thanks. :)

2 Answers

I prefer LoRAs 100% of the time. Embeddings, at least for me, only vaguely grasp the concept they are trained on and don't really mesh well when using multiple at once. LoRAs on the other hand, despite having a larger file size, are much more versatile with what you can do with them and seem to replicate their training better. They also are easier to train IMO.

LoRAs will almost always be better simply due to how the training works.

In more layman terms, TI (Embeddings) trains a model's perception/understanding of prompts only, without actuall affecting the model's "skillset".

LoRA/Dreambooth/Finetune/Hypernetworks will all affect the model's actual "skillset".

Your answer