TALECRAFTERS
GLOSSARY · MODELS · CONTROL

LORA

A small trained adapter that teaches a large model one style, character or object without retraining the model itself.

ALSO CALLEDLow-Rank Adaptation

A LoRA is a few megabytes that sit alongside a model of many gigabytes and steer it. Cheap to train, easy to version, trivial to remove. For a studio this is the right granularity: one per brand style, one per recurring character, swapped per project.

The failure mode is overcooking. A LoRA trained too hard on too few images stops being a style and starts being a photocopier, reproducing the training set’s backgrounds and poses along with the thing you wanted.

QUESTIONS PEOPLE ASK

How many images do you need to train a LoRA?

For a style, a few dozen varied examples. For a character, twenty to forty clean stills at varied angles under even light. Quality and variety matter more than quantity.

LoRA or full fine-tuning?

LoRA first, every time. Escalate to a full fine-tune only when a measured failure rate at your actual volume justifies the maintenance.

THEORY IS FREE

Knowing the word is the cheap part.

Running it on a deadline, at volume, without burning the budget is the expensive part. That is the bit we do.

BRIEF US