LoRA

LoRA is the cheap way to teach a model something new without retraining it. It trains on fifteen to twenty images and produces a small file of a few hundred megabytes.

Its three most common uses: teaching your own face (personal avatars), teaching a product (corporate imagery), teaching an illustration style (brand consistency).

The variety of training images determines the result. If they are all from one angle, the model learns that angle rather than the face, and likeness collapses at other angles.