Filter analyzing and modifying local pixel neighborhoods — median, Gaussian, bilateral. Works directly in image space, not frequency domains like FFT.
On set or in the edit, you often need quick, local image corrections — not global effects across the entire frame. This is where spatial filters come in: they examine each pixel and its immediate neighborhood, calculate a new value based on this small region, and write the result back. This happens in parallel for every point in the image, which is why the performance remains practical even at high resolutions.
The classic types: A Median filter sorts the pixel values in the neighborhood and takes the middle one — brutally effective against noise and single artifacts, without blurring edges like a Gaussian would. This is exactly what you need if your DNG sequence has sensor noise but you want to retain detail sharpness. A Gaussian Blur is the opposite: uniform weighting with falloff — it creates the classic soft blur for motion blur compositing or as a basis for local color grading. The Bilateral filter is the compromise: it smooths color gradients but respects edge transitions. If you want to smooth a face without it looking like plasticine, Bilateral is your first choice.
In everyday compositing, you use spatial filters for denoise operations before keying, for local brightness homogenization on green screens with uneven lighting, or to smooth out individual pixel errors in CGI renders before layering them. Spatial filters also work when creating mattes — Dilate, Erode: they expand or shrink white areas based on pixel neighborhoods. This is faster and more local than global transformations and gives you more precise control over problematic areas.
Important: Spatial filters operate directly in image space — not in the frequency domain like FFT-based filtering. This makes them more intuitive to control (radius instead of frequency cutoff), but also less flexible with complex artifact patterns. Larger filter kernels cost exponentially more in performance — a 5x5 matrix is still cheap, but 21x21 starts to strain. Pro tip: Separable filters (first horizontal, then vertical) for Gaussian and Median dramatically reduce complexity.