Practical Error Estimation for Denoised Monte Carlo Image Synthesis

𝑠 2 𝑖 (𝜆 𝑖 + 1) 32spp RelSE 𝑠 2 𝑖 (𝜆 𝑖 + 1) 512spp RelSE 50 90 99 99.9 10 3 logRelSE RelSE s 2 i ( i + 1) s 2 i 2 1, i Figure 1: An example demonstrating that our error estimation framework for Monte Carlo denoised images reliably estimates an image's ground truth relative squared error (RelSE) distribution at different average sample counts (spp). We show that the RelSE of a denoised pixel 𝑖 follows a scaled noncentral chi-squared distribution, with scale 𝑠 2 𝑖 and noncentrality 𝜆 𝑖 . To estimate these parameters, we hierarchically aggregate noisy per pixel estimates of error and variance. Knowing 𝑠 2 𝑖 and 𝜆 𝑖 , we can accurately estimate the image's error distribution, visualized per pixel (middle) and as the upper half of a logistic-logarithmic percentile plot (rightmost), leading to a robust stopping criterion for denoised Monte Carlo image synthesis.