Arbitrary Image Rescaling

Arbitrary image rescaling aims to develop algorithms that can resize images to any desired scale, unlike traditional methods limited to fixed factors. Current research heavily utilizes invertible neural networks (INNs), often incorporating techniques like dual latent variables and position-aware scale encoding to improve the accuracy and perceptual quality of both upscaling and downscaling, even handling asymmetric scaling. This work is significant because it addresses the limitations of fixed-scale methods, improving image quality in various applications such as social media and image editing, where flexible resizing is crucial. The focus is on achieving bidirectional rescaling (upscaling and downscaling) with high fidelity and robustness, even across repeated rescaling cycles.