Denoising

What the filter is actually doing

A denoiser is not a better renderer. It is a classifier: given a noisy estimate of a pixel's colour, it tries to decide which variation is signal — a real edge, a genuine texture — and which is noise that a longer render would have averaged away. The distinction is genuinely hard, and the denoiser will sometimes get it wrong in both directions.

Most production denoisers are convolutional or transformer networks trained on paired images: the same scene rendered at low sample counts alongside a nearly-converged version. During training, the network learns which clues correlate with real structure. The primary clue is usually not the colour channel at all. Auxiliary data — surface normals, albedo, depth — gives the filter a geometry-aware scaffold. Without those passes the denoiser is reading noise and guessing; with them, it can respect object boundaries and avoid blurring across edges it was never shown in the colour data.

The cost is a new category of artefact. Fine surface detail — scratched metal, close-up fabric — lives at the spatial frequency where noise and texture look similar. The filter, unsure, erases both and replaces them with smooth, posterised regions. Specular highlights acquire a ghosted smearing on moving frames because the network evaluates each frame independently; the highlight does not travel the same path twice. And any feature smaller than the network's receptive field — the tiny bright catch-light, a thin wire against sky — risks being classified as a firefly and suppressed.

The right question to ask of a denoised image is not "does it look clean?" but "is the detail that remains real, or is it artefact the network invented?" A region the denoiser smoothed over will look confident and wrong rather than noisy and honest. Noisy is at least informative — it tells you the budget ran out. Smooth tells you nothing about whether the answer underneath was correct.

Denoise early in a sequence and the temporal instability compounds; denoise at a high enough sample count that the filter is trimming genuine residual noise rather than reconstructing a missing signal, and the result earns its cleanliness.