We examined how well we can recover surface-reflectance properties from shading patterns under changes in surface shape. The stimulus we used was a square surface modulated in depth by a low-pass-filtered random field and rendered by the Phong illumination model [Commun. ACM 18, 311 (1975)]. Two different surface images (target and match) were presented side by side, with either the viewing direction or the surface-normal direction rotating around the horizontal axis. The target shape was manipulated by changing the spatial spectrum, and the target reflectance was manipulated by changing the diffuse-reflection coefficient and the specular-reflection exponent (shininess) of the Phong model. The shape parameters of the match stimulus were fixed, but its reflectance parameters were under the control of subjects, who had to make the apparent reflectance of the two surfaces as similar as possible. The results showed that the constant error (difference between simulated and matched values) was large except when the two surfaces had the same shape parameters or when they differed only in scale. The pattern of the constant errors and response variabilities suggests that the judgments of the subjects were based on the similarity of the luminance histogram of the surface image. Our results demonstrate a limitation of surface-reflectance constancy for changes in shape and the importance of image-based information in reflectance judgments. The results are discussed in relation to previous studies that showed effects of spatial layout on surface-reflectance perception.