We investigated visual estimation of surface roughness of 3-d textures under various lighting conditions. To construct a texture, first we made a jittered planar grid. Next, vertices were chosen above each grid intersection with a random height. Vertices were joined to create a triangulated surface. The variance of the height distribution was our measure of roughness. These surfaces, with Lambertian (matte) reflectance, were rendered for binocular viewing (viewing distance: 70 cm) under one of three punctate illuminants located 20, 30 or 40 deg to the right of the surface normal. In Experiment 1, the surface patch was viewed in isolation through an aperture. Observers viewed two surface patches in succession, each rendered under a different punctate illuminant, and indicated which appeared to be rougher (2 IFC). Using interleaved staircases, for each pair of lighting conditions, we determined the roughness of a surface under one illuminant that appeared equally rough as a given surface under the other illuminant. Participants showed a strong bias for perceiving a surface to be rougher as the illuminant moved further away from the surface normal. Thus, observers were not roughness constant. In a second experiment, we removed the aperture and added both matte and specular objects (and their cast shadows) to the scene to provide additional cues to the direction of the illuminant. Roughness discriminations were performed with and without the added objects. The additional illuminant cues improved roughness constancy, but only slightly. In addition, the results displayed approximate transitivity: if surface A in 20 deg illumination appeared equally rough as B in 30 deg illumination, and likewise B in 30 deg and C in 40 deg, then A and C appeared approximately equally rough in their respective illuminants. This is the first study to directly test the visual system's ability to maintain a constant estimate of roughness under changing illumination conditions.