Updating every weight in a large model needs serious hardware and produces a full-sized copy for each variant. LoRA trains a small set of additional parameters alongside the frozen original, which captures most of the adaptation at a tiny fraction of the cost, and yields an adapter that can be swapped in and out at generation time.
In practice that changes what is worth trying. Adapters are cheap enough to keep several, stack them, and throw away the ones that did not work. The artwork on this site was produced that way: a look trained once, then applied consistently across every room, prop and sprite, which is a problem that prompting alone does not solve at that scale.
Related terms
The bench this belongs to
Generative AIStable Diffusion and ComfyUI wired into the place where your content actually gets made, with a model fine-tuned so everything comes out looking like you.
