Reproduce “Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts” 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
Hang Guo, Tao Dai, Yuanchao Bai, Bin Chen, Xudong Ren, Zexuan Zhu, Shu-Tao Xia
AdaptIR adapts pretrained restoration models to degradations by training an add-on while keeping pretrained weights fixed. Its add-on combines local, frequency-domain, and channel-gating branches with task-adjusted outputs. On Urban100 for ×4 downsampling plus noise, it reaches 22.88 dB PSNR, versus 22.77 for a multi-branch baseline; the result is not a win on every metric or evidence of zero-shot restoration.
Key insights
- Homogeneous representations learned by existing PETL methods lead to unstable performance when applied to diverse image restoration tasks, especially for hybrid degradations.
- Learning heterogeneous, degradation-specific representations, by leveraging orthogonal feature modeling (local, global, channel), is crucial for robust image restoration.
- Inserting parameter-efficient modules parallel to the MLP layers of a frozen transformer backbone optimally preserves pre-trained knowledge while allowing effective adaptation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.