ControlNet.

Definition

ControlNet is an add-on to an image diffusion model that forces the generated picture to respect a structural reference — a depth map, an edge drawing, a pose skeleton, a segmentation map, a scribble. It clones the base model's encoder into a trainable copy connected by zero-initialised layers, so the original weights stay frozen and the control starts as a no-op instead of degrading the model. Practically it separates two decisions that a text prompt fuses together: composition (what goes where, set by the reference image) and style (how it looks, set by the prompt). That is what makes generative image models usable in design work — you can iterate on look while the layout, framing and proportions hold.