Apply a convolution mask to an image — pixels multiplied and summed with neighbors. Core operation in compositing and digital grading.
Convolve
You apply a convolution mask when you combine each pixel of an image with its immediate neighbors—a mathematical operation that is fundamental in digital image processing. The mask (a small kernel grid, usually 3x3 or 5x5) is slid across the entire image. For each position, each pixel under the mask is multiplied by a coefficient, all products are summed, and the result is written as a new pixel value. This sounds abstract, but you recognize the result immediately: blur, edge enhancement, noise reduction—all convolutions.
In practical compositing, you constantly need convolutions. A blur node in Nuke is a convolution—it takes each pixel and mixes it with its neighbors in a weighted pattern. A sharpen filter works the same way: high weights for the central pixel, negative values for the neighbors. Edge detection (Sobel, Laplacian) is also pure convolution—special kernels highlight brightness edges. You quickly see: the kernel defines everything. A Gaussian blur is simply a kernel whose coefficients map a Gaussian distribution.
Grades and color correction use convolutions for spatial effects. A localized blur before color adjustment, grain reduction through adaptive convolutions, or spatial frequency separation—these are all convolutions. Some grading software offers a custom kernel editor so you can build exotic effects. In the Red workflow or DaVinci Resolve, you sometimes apply convolutions unnoticed: every spatial blur, every denoise operation that considers pixel neighbors, is a convolution.
Performance note: Large kernels (11x11 or larger) can be expensive—they sample 121+ neighbor pixels per position. GPU acceleration makes it faster. Separable kernels (e.g., horizontal convolution, then vertical) save massive computation time. On set or in post, you should know: a simple Gaussian blur is practically free. You must consciously set a custom kernel with edge handling (extend, wrap, mirror?), otherwise artifacts will appear at the image edges. Convolution is not new—it originates from signal processing—but in the modern VFX pipeline, it is omnipresent and often invisible.