Local Weighted Regression

Local Weighted Regression is a noise reducing filter based on the local regression model. Denoising is done by obtaining coefficients for nearby pixels within a small (local) window to reconstruct the pixel color. To preserve contrasted edges, the bandwidth parameter is used.

No denoiser

Samples: 4, Radius: 4, Bandwidth: 0.2

Samples: 4, Radius: 4, Bandwidth: 0.5

Download the above scene (mb.zip)

Samples

The Samples parameter defines the number of samples (iterations) taken to calculate the color of each pixel. Greater values produce better results but require more time for denoising.

Filter Radius

The Filter Radius parameter controls the area or window within which similar neighboring pixels will be found.

The parameter defines the area radius in pixels, and supports values from 1 to 10. Greater values give a wider area for sampling nearby pixel colors, which yields a greater effect but requires more resources and time to calculate the resulting pixel color.

Bandwidth

The Bandwidth parameter controls the bandwidth of the filter, or how much of the data is used to calculate the resulting pixel color. Smaller values give less noise but may result in loss of image details.