Scene 05 · Train
Custom Model Training
Custom models (LoRA / qLoRA) trained on your subject, style, or motion, so generated work stays on-model across an entire campaign — and into production.

A prompt gets you close. A trained model gets you the same face, the same product, the same house style every single time.
We train on your references, images or video, and hand back the weights. The training set stays here, and the model is yours to put into a real production pipeline.
Questions
- How much reference material do you need?
- Less than people expect. Often a few dozen good frames. We will tell you if a set is too thin before we start.
- Can the model go into our own pipeline?
- Yes. You keep the weights and run them wherever you need — or on our private compute if the work has to stay inside a perimeter.
- Is my training data used for anything else?
- Never. It trains your model and nothing else.