Models and architecture

    LoRA

    Also called Low-Rank Adaptation, Adapter.

    A LoRA is a small add-on file that adjusts a large model's behaviour — teaching it a specific character, product or style — without retraining or replacing the model itself.

    Rather than changing the model's billions of parameters, the technique inserts small matrices alongside them and trains only those. The result is a file that can be megabytes instead of gigabytes, trained from a modest set of example images, and loaded on top of a base model when you want its behaviour and left off when you do not.

    That makes it the practical answer to identity at scale. A character you will use for a hundred clips is worth teaching once; after that, the subject is available from a prompt keyword rather than assembled from reference images every single generation.

    Two constraints matter. A LoRA is bound to the base model it was trained against and does not transfer to a different one. And several loaded at once compete — stack a character, a style and a lighting adapter and the result is rarely the sum of the three.

    In practice

    • Small file, trained on a small set, loaded on top of a base model.
    • Tied to its base model — a new model version means retraining.
    • Stacked adapters interfere; weight them down or use fewer.

    The mistake to avoid

    Training one on twenty photos from the same shoot. The adapter learns that day's lighting and background as part of the subject and reproduces them everywhere.

    Related terms

    The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.