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.
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.