Arbitrary Scale Image Super Resolution

Arbitrary-scale image super-resolution (ASSR) aims to upscale low-resolution images to any desired resolution, overcoming the limitations of fixed-scale methods. Current research heavily utilizes implicit neural representations (INRs), often employing novel architectures like Gaussian splatting or mixtures of experts to improve efficiency and reconstruction quality while addressing challenges such as blurry outputs and computational cost. These advancements are significant for various applications requiring flexible image scaling, including image editing, forensics, and computer vision tasks that benefit from high-resolution inputs at arbitrary scales.