Mathematical matrix for digital filtering — controls blur, sharpen, motion effects in post. Larger kernel = heavier computation, more precision.
In digital image processing, you constantly work with kernels without using the term – every time you apply a blur filter or adjust sharpening. The kernel is the mathematical template behind it: a small matrix of numerical values that is overlaid on each pixel of an image. The kernel value is multiplied by the neighboring pixels, the results are added – and the new pixel value is ready. The size and values determine what happens.
On set or in the VFX suite, you encounter this when working with convolution filters. A 3x3 kernel is the smallest sensible unit – quick to calculate, but with limited effect. A 5x5 or 7x7 kernel becomes more computationally intensive, but delivers more subtle results. Motion blur is different: here you often need larger kernels to credibly simulate the direction and intensity of movement. An incorrectly sized kernel leads to visible artifacts – kernels that are too small for strong effects, and you see individual blocks; too large, and the calculation takes forever without much visual improvement.
In compositing software like Nuke or After Effects, you see this concretely when you can tune filter parameters: The radius of a blur filter is directly linked to the kernel size. A Gaussian blur with a radius of 10 creates a larger internal kernel than a radius of 2. For sharpening (Unsharp Mask), you use a kernel that amplifies high frequencies – typically a negatively weighted neighbor mix that accentuates edges. The art lies in choosing the right kernel that fulfills your artistic intention without crashing the render farm.
Practically: If your VFX supervisor criticizes render times, first look at the active filters. An unnecessarily large kernel for chromakeying or rotoscope feathering costs more processing time than it provides visible quality gain. Some software allows you to define custom kernels – but then you need a solid understanding of matrix mathematics to avoid working randomly. For everyday use: Standard kernels in professional software are optimized; only for very specific requirements (e.g., artistic distortion effects) is the detailed work worthwhile.