ControlNet lets you use a start image to control your generation in specific ways — much more powerfully than a plain start image.
A start image on its own (image-to-image) just initialises the diffusion process instead of starting from random pixels; it doesn’t actively guide the result. ControlNet models, by contrast, guide the generation, a bit like a visual prompt. For example:
By default, ControlNet runs for 100% of the chosen steps. Lowering this stops ControlNet early and lets the normal diffusion process finish the image. This is useful when you want ControlNet to set the initial direction, then let the model take over for the rest.
By default, your start image is pre-processed for certain ControlNet models — e.g. converted to canny edges for “Canny Edge to Image”, or to a depth map for “Depth to Image”. If you already have a depth map or edge map, turn on Skip Preprocessing to use your start image directly.