What do the low and high thresholds actually do?

Canny uses hysteresis. Pixels with a gradient above the high threshold are definite edges; pixels above the low threshold are kept only if connected to a definite edge. Raising the high threshold drops weak edges, while lowering the low threshold lets more faint detail through.

What thresholds should I start with?

A common default is a low of 100 and a high of 200. For busy photos where you want only major contours, raise both (for example 150/300). For line art or sketches where you want every line, lower them (for example 50/150).

What is aperture size?

Aperture size is the Sobel kernel used to compute gradients, usually 3, 5, or 7. A size of 3 is the standard and captures fine detail; larger apertures smooth the image first, producing thicker, fewer edges that suit noisy or high-resolution sources.

Should I enable L2 gradient?

L2 gradient uses a more accurate Euclidean gradient magnitude instead of the faster L1 approximation. It produces slightly cleaner, more precise edges at a small speed cost. Enable it for final maps where edge accuracy matters; the default L1 is fine for quick iteration.

Canny Edge Preprocessing Guide

Canny edge preprocessing guide

Canny edge detection is the most-used ControlNet preprocessor: it turns a source image into a clean black-and-white edge map that constrains the structure of the generated image. Its output is almost entirely controlled by two numbers — the low and high thresholds — plus the aperture size and the L2 gradient toggle. Get the thresholds wrong and you either lose the composition or trace every speck of noise.

How it works

Canny finds edges in stages and uses hysteresis to decide which to keep. Any pixel whose gradient exceeds the high threshold is a definite edge. Any pixel above the low threshold is kept only if it connects to a definite edge — this links genuine contours while rejecting isolated noise. The aperture size sets the Sobel kernel for the gradient computation (3 is standard; larger smooths first), and L2 gradient swaps the fast L1 approximation for a more accurate magnitude. The map this produces is what ControlNet locks the generation’s structure onto.

Tips for the right edge density

Match thresholds to source. Line art wants low thresholds (catch every line); busy photos want high thresholds (keep only major contours).
Widen the gap for more detail. A bigger spread between low and high (say 50/200) preserves more connected faint edges than a narrow one.
Use aperture 5 for noisy images. Larger apertures smooth before detecting, giving cleaner maps from grainy or high-resolution photos.
Save L2 gradient for finals. It is slightly slower but produces crisper, more accurate edges when the map quality matters.