Reproduce “Self-Asymmetric Invertible Network for Compression-Aware Image Rescaling” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jinhai Yang, Mengxi Guo, Shijie Zhao, Junlin Li, Li Zhang

A new Self-Asymmetric Invertible Network (SAIN) from Bytedance enables robust image rescaling that accounts for lossy compression artifacts. The model achieves PSNR gains of 1.33-3.59 dB over competing methods for x2 upscaling from JPEG-compressed inputs, while also producing high-quality downscaled images for efficient storage and transmission.

Key insights

  • Standard invertible networks are highly vulnerable to real-world lossy compression, which introduces significant distribution shifts and degrades reconstruction quality.
  • The 'lost information' (high-frequency components) in compression-aware invertible networks follows a multi-modal distribution, necessitating a more flexible model than a simple Gaussian.
  • Integrating a differentiable compression simulator and an asymmetric design within a single invertible network enables robust learning of compression effects for superior upscaling, while maintaining high-quality downscaling outputs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.