Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for rendering an image of a scene affected by multiple light sources. In one aspect, a method includes subdividing the scene into cells; sampling light source —point pairs; for each pair, determining a contribution value of the light source to the point; for each cell and each light source: determining a maximum contribution value of the contribution values for the light source to the color of the points that are in the cell, and determining, based on the maximum contribution value, a significance value that is a measure of an estimated importance of the light source in rendering a portion of the image corresponding to the cell; and rendering the image of the scene by sampling light sources having a higher significance value more often than light sources having a lower significance value.