Unbiased rendering


Within the field of computer graphics, unbiased rendering refers to any rendering technique that does not introduce systematic error, or bias, into the radiance approximation. The term refers to statistical bias, not the broader meaning of subjective bias. Because of this, an unbiased rendering technique can produce a reference image to compare against renders that use other techniques. Path tracing and its derivatives can be unbiased, whereas ray tracing was originally biased.

Mathematical definition

Mathematically speaking, the expected value of an unbiased estimator is the population mean, regardless of the number of observations. The error found in a render produced by an unbiased rendering technique is due to random statistical variance, which manifests as high-frequency noise. Variance is reduced by for data, meaning that four times as many data are needed to halve the standard deviation of the error; this makes unbiased rendering techniques less attractive for realtime or interactive applications. This means that an image produced by an unbiased renderer that appears noiseless and smooth is probabilistically correct.
A biased rendering method is not necessarily wrong, and can still produce images close to those given by the rendering equation if the estimator is consistent. These methods, however, introduce a certain bias error in efforts to reduce the variance. Often biased rendering is optimized to compute faster at the cost of accuracy.

Caustics example

It is important to note that an unbiased technique can not consider all possible paths, and may not select ideal paths for a given render. Path tracing, an unbiased approach at its core, cannot consistently handle caustics generated from a point light source, as it is highly unlikely to randomly generate the singular path that directly reflects into the point.
Progressive photon mapping, a biased rendering technique, can actually handle caustics quite well. Although biased, PPM is provably consistent, meaning that as the number of samples goes to infinity, the bias error goes to zero, and the probability that the estimate is correct reaches one.

List of unbiased rendering methods

*