Small matrix of weight values applied to each pixel — 3×3, 5×5 or larger. Determines the image operation: Gaussian blur, Sobel edge, sharpening.
On the monitor sits a 3x3 matrix of numbers — that's your tool for pixel manipulation. You place this matrix over each individual pixel of your image, multiply the neighboring pixels by their corresponding weight values, and sum up the result. The result becomes the new pixel value. You repeat this process for every position in the image — and your image operation is complete. This is convolution, and the convolution kernel is the matrix itself.
The size and the values determine everything. A symmetrical 3x3 matrix with all ones, divided by 9, creates a blur effect — each pixel becomes the average of its neighbors. If you increase the weights in the center, you amplify the original information and create blur with presence. The famous Gaussian blur essentially boils down to a Gaussian distribution as the kernel — larger matrices (5x5, 7x7) with decreasing weights towards the outside. If you want to detect edges, you use edge detection kernels like Sobel or Roberts — there, neighboring pixels have different signs, which compresses transitions. Sharpening works similarly: a high central value (e.g., +5), negative neighbors — this enhances contrasts and adds definition.
On set, this is secondary, but in post — in Nuke, After Effects, or during the grading pipeline — convolution kernels are omnipresent. You won't manually write a matrix every time; the software has ready-made filter libraries with optimized kernels. But if you need custom looks or want to understand what's happening under the hood, you need to internalize the principle. A 5x5 matrix costs more processing power than a 3x3 — this is relevant for VFX-heavy shots. Larger kernels yield smoother results, but also more latency. Some DoPs and VFX Supervisors build up blur stacks iteratively, several weak convolutions instead of one strong one — this often looks more natural and is also more performance-friendly. An important point: kernel operations are separable for certain filters (like Gauss). You can first convolve horizontally, then vertically — this saves massive processing power and is therefore standard in real production software.
Remember: what the kernel doesn't "see" (the edges of the image) must be extrapolated — padding strategies like Mirror, Wrap, or Constant Black change the result here. In the final composite, you'll notice this as an edge artifact if it's configured incorrectly.