In natural image denoising, the task is to estimate a clean version of a given noisy image, using prior knowledge on the statistics of natural images. The problem has been studied intensively with impressive progress achieved in recent years. However, it seems that image denoising has reached a plateau, with new algorithms improving over previous ones by only fractional dB values. A key question is thus: How much more can current methods be improved? In this talk we’ll discuss optimal natural image denoising and its fundamental lower bounds. In particular, we’ll show that at moderate noise levels, current state-of-theart denoising algorithms, that use a fixed small support window around each denoised pixel, are approaching optimality and cannot be further improved beyond fractional dB values. Joint work with Anat Levin.