Texturing is a ubiquitous technique used to enrich surface appearance with fine detail. While standard filtering approaches, such as mipmapping or summed area tables, are available for rendering diffuse reflectance textures at different levels of detail, no widely accepted filtering solution exists for multiresolution rendering of surfaces with fine specular normal maps. The current state of the art offers accurate filtering solutions for specular reflection at the cost of very high memory requirements and expensive 4D queries. We propose a novel normal map filtering solution for specular surfaces which supports data pre-filtering, and with an valuation speed that is roughly independent of the filtering footprint size. Its memory usage and evaluation speed are significantly more favorable than for existing methods. Our solution is based on high-resolution binning in the half-vector domain, which allows us to quickly build a very memory efficient data structure.