CUF: Continuous Upsampling Filters

Figure 1. We report (left) memory and FLOPs for upsampling an 256 × 256 image by different scale factors (2×,3×,4×) using integer scale upsamplers (Sub-Pixel Convolution [31] and CUF-instantiated) and arbitrary-scale upsamplers (Meta-SR [16], LIIF [6], LTE [19] and CUF), but the same encoder backbone (EDSR-baseline [21]); and (right) the relationship between each upsampler FLOPS and PSNR performance on DIV2k dataset. Our arbitrary scale model is significantly lighter than other methods in the same class (i.e. continuous super-res). Further, when instantiated for integer scale factors, our upsampler is even-more efficient than sub-pixel convolutions [31].