Denoise

Denoise

Noise is the enemy of every render. When a renderer does not have enough samples, the image is covered in a grainy pattern called Monte Carlo noise. This happens most often in dark areas, in areas with complex indirect lighting, and in motion blurred objects. Denoising is the process of removing this noise while preserving the actual detail in the image. It is an essential step in any VFX pipeline.

Traditional denoising works by blurring the image while trying to preserve edges. A simple blur removes noise but also removes detail, making the image look soft. Better denoising algorithms use edge detection to avoid blurring across edges. The bilateral filter is a classic example. It blurs pixels that are similar in color and close in position, but does not blur across strong edges where colors change sharply.

More advanced denoising uses information from the render itself. Modern renderers can output additional data like surface normals, world positions, and albedo alongside the beauty pass. Denoisers use this extra data to guide the filtering. If two pixels have very different normals, they are probably on different surfaces and should not be blurred together. This gives much better results than blind denoising.

Temporal denoising uses information from neighboring frames. If a pixel is noisy in one frame but clean in the next, you can average them together. This works well for static parts of the image but can cause ghosting on moving objects. Good temporal denoisers use motion vectors to track where each pixel moves between frames, so they only blend pixels that correspond to the same surface.

The best approach is to render enough samples that the noise is minimal and then use a light denoise to clean up the rest. Relying on heavy denoising to fix a very noisy render will always produce artifacts. The goal is to find the sweet spot where you render just enough samples to capture the detail, and let the denoiser handle the remaining noise without destroying the image quality.