Models and architecture

    Fine-tuning

    Fine-tuning continues training an existing model on your own examples so it produces your subject or style by default, rather than on request.

    You are not building a model from nothing — that takes a data centre. You are taking one that already knows what pictures and motion look like and nudging its weights towards a narrower target using a comparatively tiny set of examples.

    The dataset decides everything. A few dozen varied, well-labelled examples beat hundreds of near-identical ones, because the model learns whatever is consistent across them — and if every photo was shot in the same room, the room is part of what it learns. Variety in angle, light and context is what separates a subject from a memorised snapshot.

    Overfitting is the standard failure and it looks deceptively like success. The model reproduces your examples beautifully and has lost the flexibility to place your subject anywhere new, which is exactly the capability you wanted.

    In practice

    • Vary the examples deliberately: angles, distances, lighting, backgrounds.
    • Label consistently — inconsistent captions teach inconsistent behaviour.
    • Adapter methods like LoRA are the light-touch version of the same idea and are usually enough.

    The mistake to avoid

    Fine-tuning for something a reference image would have handled. Training is a commitment; references are a per-job decision you can change tomorrow.

    Related terms

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