LoRA: Training a Model on a Particular Likeness

A LoRA is a small addition to an existing model that teaches it one specific thing: a face, a style, an object. Retraining a whole model for one person is impossible; training an addition takes an hour on a home graphics card.

How it works

Instead of altering billions of model parameters, a compact set of corrections to them is trained. The file comes out at tens of megabytes rather than gigabytes, attaches to the base model on demand and detaches the same way.

How many images are needed

For a face, fifteen to forty frames is usually enough. More is not better: variety matters more than quantity. Thirty images with different lighting and angles beat two hundred from a single session.

• Different angles: full face, three-quarter, profile
• Different light: studio, daylight, evening
• Different framings: close-up, waist-up
• Consistent quality: one blurred frame in the set spoils the whole sample

Overtraining

The commonest beginner mistake is training too long. An overtrained addition stops being an addition: it drags along the background, clothing and poses from its training frames, and the likeness becomes impossible to place in a new scene. An undertrained one merely exerts weaker influence — that is fixable, overtraining is not.

The legal side

Training an addition on someone else's photographs means building a tool for reproducing their likeness. Even if the images came from an open profile, the result falls under the same restrictions as any work with a person's image: consent is required. Nothing prevents it technically; everything prevents it legally.

This material is intended for adults over 18 and is informational in nature.