HED Edge Detection

HED stands for Holistically-Nested Edge Detection, a deep learning architecture developed by Saining Xie and Zhuowen Tu that uses fully convolutional networks for image boundary extraction. In generative AI workflows, HED is the standard preprocessor for Soft Edge ControlNet pipelines, delivering organic, sketch-like contour maps rather than rigid geometric vectors.

HED produces soft, continuous grayscale lines with varied pen-pressure gradients that mimic hand-drawn artist sketches, whereas Canny produces hard, binary, single-pixel lines. Reach for HED when preserving organic silhouettes, drapery, and human anatomy while giving downstream diffusion models freedom over fine texture; reach for Canny when you need rigid architectural wireframes or precise CAD-like borders.

Safe mode filters out faint background grayscale textures and subtle image noise from the source photo before generating the final boundary map. Enable safe mode when feeding downstream Soft Edge ControlNet models to prevent ghosting artifacts, color bleed, or texture hallucinations from leaking into your final render.

Scribble mode binarizes and thickens the extracted contours to simulate bold felt-pen doodles and hand-drawn ink sketches. Use scribble mode when pairing with Scribble ControlNet checkpoints to achieve punchy graphic outlines, but leave it disabled for standard Soft Edge models to preserve delicate tonal gradients.

HED excels at broad silhouettes, body poses, and organic flow, but it struggles with pixel-perfect typography, intricate CAD blueprints, and high-frequency micro-textures like fine hair strands or tight fabric weaves. For technical vector alignment or micro-geometry, use Canny or Lineart preprocessors instead.