Reference Image Rendered Image Error Threshold: 0.25 Error Threshold: 0.125

tua l E rro r Im age s Figure 1: Our error estimation framework takes into account error due to noise and bias in progressive photon mapping. The reference image is rendered using more than one billion photons, and the rendered image is the result using 15M photons. The noise/bias ratio shows the areas of the image dominated by noise (in green) or bias (in red). The bounded pixels images show the results corresponding to a desired confidence in the estimated error value (pixels with bounded error are shown yellow): a higher confidence (90%) provides a conservative per pixel error estimate, and lower confidence (50%) is useful to estimate the average error. One application of our error estimation framework is automatic rendering termination with a user-specified error threshold (bottom row: the images show color-coded actual error). Our framework estimates error without any input from a reference image. We present an error estimation framework for progressive photo