Résumé : In this paper we present a novel image based algorithm to render visually plausible anti-aliased soft shadows in a robust and efficient manner. To achieve both high visual quality and high performance, it employs an accurate shadow map filtering method which guarantees smooth penumbrae and high quality anisotropic anti-aliasing of the sharp transitions. Unlike approaches based on pre-filtering approximations, our approach does not suffer from light bleeding or losing contact shadows. Discretization artefacts are avoided by creating virtual shadow maps on the fly according to a novel shadow map resolution prediction model. This model takes into account the screen space frequency of the penumbrae via a perceptual metric which has been directly established from an appropriate user study. Consequently, our algorithm always generates shadow maps with minimal resolutions enabling high performance while guarantying high quality. Thanks to this perceptual model, our algorithm can sometimes be faster at rendering soft shadows than hard shadows. It can render game-like scenes at very high frame rates, and extremely large and complex scenes such as CAD models at interactive rates. In addition, our algorithm is highly scalable, and the quality versus performance trade-off can be easily tweaked.