Self-Asymmetric Invertible Network for Compression-Aware Image Rescaling

High-resolution (HR) images are usually downscaled to low-resolution (LR) ones for better display and afterward upscaled back the original size recover details. Recent work in image rescaling formulates downscaling upscaling as a unified task learns bijective mapping between HR LR via invertible networks. However, real-world applications (e.g., social media), most compressed transmission. Lossy compression will lead irreversible information loss on images, hence damaging inverse procedure degrading reconstruction accuracy. In this paper, we propose Self-Asymmetric Invertible Network (SAIN) compression-aware rescaling. To tackle distribution shift, first develop an end-to-end asymmetric framework with two separate mappings high-quality respectively. Then, based empirical analysis of framework, model lost (including compression) using isotropic Gaussian mixtures Enhanced Block derive high-quality/compressed one forward pass. Besides, design set losses regularize learned enhance invertibility. Extensive experiments demonstrate consistent improvements SAIN across various datasets terms both quantitative qualitative evaluation under standard formats (i.e., JPEG WebP). Code is available at https://github.com/yang-jin-hai/SAIN.