In our recent work [6], we proposed an algorithm for image upsampling based on alternation of two procedures: spatially adaptive Þltering in image domain and projection on the observationconstrained subspace in a wavelet domain. The nonlocal Block-Matching 3-D (BM3D) Þlter was used to suppress ringing and reconstruct missing detail coefÞcients. Here we generalize this method in two aspects. First, we reformulate observation model and reconstruction algorithm from wavelets to a general class of scaling transforms, and second, we extend applicability from single image upsampling to image and video super-resolution. Experimental results demonstrate signiÞcant improvement over the upsampling method [6] and an overall performance on the level of the best existing super-resolution methods. 1.