Typing a prompt and hoping for the right layout is a gamble. Pros skip the gamble: they hand the generator a structure map — your exact lines, a pose, a shape — and force it to obey while it fills in the colors and style. Draw some lines below and flip control ON to watch it stay inside them.
Sketch anything in the first box — a house, a face, blobs. The middle box shows the real edge map (a Sobel filter, the same "outline detector" pros feed a generator). Hit Generate: with control ON the colors stay inside your lines; flip it OFF and the generator ignores you and paints whatever. That flip is the whole lesson.
Generate once with control ON, then flip it OFF and Generate again on the same lines. Same drawing, wildly different result — because OFF throws your structure away. 🎛️
Edges are one kind of control. Here's another: a pose. Drag the dots to move the head, hands, and feet. Hit Generate and the SAME style engine fills a character that obeys your skeleton. Real tools call this pose control — you decide the stance, the AI decides the looks.
Move an arm up and Generate again — the character throws the same arm up. That's structure you control, style the AI invents. Two different control types, one obedient generator.
The structure says ⬛ square outline. But now ask for something round or spiky in the style. The generator can't fully obey both — so it compromises. Tap a request and watch the square bend toward it without ever becoming a clean circle.
A faint target silhouette is ghosted in your drawing box. Draw a closed outline around it (trace the ghost!), then hit Generate & check. If the filled region matches the target closely enough, you score. Match 3 targets to win — and you'll learn why a control outline has to be closed.
You used edges and a pose. Real ControlNet takes a whole family of conditioning images — pick the one that captures the structure you care about:
An edge detector turns any photo into clean lines. The generator keeps every outline — great for redrawing a sketch in a new style.
A stick-figure of joints. Set the exact stance of a character; the AI fills the body. Same pose, endless outfits.
A depth map fixes what's near/far; a segmentation map says "sky here, road there". Lock the whole scene's geometry.
You stopped hoping and started commanding. You fed a generator a structure map — edges and a pose — and forced it to obey while it invented the style. You felt control ON vs OFF, watched a contradiction get compromised, and matched a target layout on purpose. That's ControlNet.