Lanczos is a new anti aliasing filter implemented in the last versions of YafaRay. A rendered image can be considered as a signal that varies over space. An anti-aliasing filter is used to restrict the bandwidth of such a "signal" to approximately satisfy the sampling theorem:
http://en.wikipedia.org/wiki/Sampling_theoremRoughly speaking, it filters out rays that does not fit in the continuity and smoothness of the final image. However, an anti-aliasing filter is also a trade-off, particularly in high contrast areas. Some filters reveal more detail but they also produce more aliasing on high contrast areas such as highlights and visible area lights. This is particularly true for two filters whose functions have negative lobes, which are
Mitchell-Netravali and
Lanczos. Both are used to amplify detail, but both will amplify aliasing on high contrast areas as well, many times beyond any reasonably anti-aliasing strategy.
Lanczos function negative lobe is bigger than
Mitchell-Netravali with a final positive curve, therefore it will reveal more detail than Mitchell but it will also produce more aliasing on high contrast areas. The negative lobe means that Lanczos reverberates over the edges of the scene (objetcs, shadows, hightlights, texturing) which brings detail up.
Which one to use ?The faster one but with the poorer results is
Box, so it is recommended for tests.
Gauss is the "perfect" filter when you can reveal detail by just using a very big imagen resultion and/or you want to keep your highlights aliasing in check, for instance when you are doing studio lighting of a car with arealights and lot of mirror component. Gauss also blurs montecarlo noise a bit.
Mitchel and
Lanzos are used to reveal and amplify detail on images which have little or no high-contrast areas and therefore diffuse components prevail.
A comparison of filters with images:
https://de.wikipedia.org/wiki/Rekonstruktionsfilter.