Reproduce “Divide and Restore: A Modular Task-Decoupled Framework for Universal Image Restoration” 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

Joanna Wiekiera, Martyna Zur

DaR-Net, a modular framework from the Silesian University of Technology, decouples degradation classification from restoration tasks using a Mixture-of-Experts approach to address universal image restoration. This framework achieved competitive performance in denoising, deblurring, and overexposure correction while significantly reducing training time and computational overhead compared to monolithic universal restorers.

Key insights

  • Decoupling degradation identification from restoration using a hard-gated routing mechanism effectively prevents negative task interference and enables efficient, specialized processing.
  • Independent training of computationally efficient U-Net experts on commodity hardware drastically reduces training time and computational overhead compared to joint optimization of monolithic universal restorers.
  • A modular design, combined with continual learning and experience replay, provides robust extensibility, allowing new degradation types to be integrated without requiring full system retraining or degrading performance on existing tasks.

Summary by alphaXiv.

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