References rent her face. A LoRA buys it.
How to train a LoRA that locks her face forever
A LoRA is a small adapter trained on 12–20 photos of your character that teaches the model her face permanently. Training takes around twenty minutes and costs a few dollars on hosted tools. After it, you stop attaching references to every generation — she is simply in the model.
What a LoRA actually is
After training, her face is a word you type — not files you attach.
Reading locks nothing in. The step this lesson teaches takes minutes in the Studio.
Build her training setCan I train a LoRA from AI-generated photos?
Yes — that is the standard workflow for synthetic characters: anchor set first (lesson 1.2), master set from it, LoRA from the master set. The model learns the face regardless of where the photos came from.
How do I know the LoRA is good enough?
The 6-photo test: three angles, two expressions, fresh scenes. If all six read as the same person to someone who has never seen her, ship it.
Does a LoRA help with video too?
Indirectly but decisively: video starts from reference images, and a LoRA gives you unlimited on-model references in any pose or outfit the scene needs.