FraIR: Fourier Recomposition Adapter for Image Restoration
Authors: Cansu Korkmaz, Nancy Mehta, Radu Timofte
Venue: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Jun. 2026
Overview
FraIR is a Fourier-based recomposition adapter for parameter-efficient transfer learning in image restoration. It applies a 1D Fourier transform to decompose token features into frequency components, performs low-rank adaptation via spectral projections with learnable reweighting, and reconstructs the adapted signal with an inverse transform gated by task-specific modulation. Integrated as plug-and-play modules within Transformer layers, FraIR is reparameterizable for zero-latency inference and adds less than 0.5% additional parameters.
Key Contributions
- Spectral-domain adapter that performs low-rank adaptation on Fourier-decomposed token features
- Reparameterizable design achieving zero-latency inference with under 0.5% additional parameters
- Outperforms prior parameter-efficient transfer learning (PETL) methods across denoising, deraining, super-resolution, and hybrid-degradation benchmarks