From one good frame to a trained model

What you are actually building

A LoRA is a small set of weights that adjusts a base model toward one subject. You invoke it by a trigger word in the prompt, and from then on the model has learned the face rather than being handed a photograph of it. That difference is why it holds at angles a reference image cannot cover — the reference only knows the view it was taken from.

The practical consequence is that everything downstream gets easier. Shots you could not attempt on the reference route — profiles, walk-toward-camera, anything in motion — become ordinary.

The training set is the whole job

Twelve to twenty frames, and the variety matters more than the count. Different angles, at least two lighting setups, at least three distances. A set of twenty near-identical frames teaches the model one pose rather than one person, and you will not discover that until you ask for something the set never showed.

Generate the set on the reference route first. Approve one frame, use it as a character reference, and produce the variety from there. This is the step people skip, and it is the reason most disappointing LoRAs are disappointing.

Include a profile and a back-of-head frame even though you will never post them. They are what stop the model inventing the parts it never saw.

Decision one: which base model

A LoRA is bound to the base model it was trained on. Moving to a different base means retraining from the same set — which is survivable if you kept the set, and expensive if you did not.

So keep the training set. Version it alongside the character sheet, in the same place, with the seed token in the folder name. The sheet plus the set is the persona; the weights are a build artefact.

Decision two: where it lives

Weights are a file, and a file has a home, a backup and an access policy. If more than one person will use the persona, decide now who can generate with it, because a LoRA that circulates is a face that can be put in frames you did not approve.

This is also the point at which the ethical guardrails stop being abstract. A trained model of an invented person is a tool. A trained model of a real person, built without their written permission, is a publicity-rights problem in most of the markets this site covers, and the training set is the evidence.

What it does not fix

Hands. Packaging text. The physics of a loaded barbell. A trained face makes the face reliable and leaves every other hard thing exactly as hard, which is why the shot library still flags product-in-hand as the frame that deserves the most attention per generation.