GSNR: Graph Smooth Null-Space Representation for Inverse Problems

A team of researchers has introduced GSNR (Graph Smooth Null-Space Representation), a novel framework that significantly improves the quality of solutions for ill-posed inverse problems in computer vision. Published on arXiv and accepted to CVPR 2026, the work addresses a fundamental challenge: when reconstructing images from incomplete or degraded data (like a blurry photo), there are infinitely many possible solutions that fit the observed measurements. Traditional methods use priors like sparsity but fail to constrain the 'null-space'—the component of the image that is invisible to the sensors. GSNR innovates by using graph Laplacians to impose smoothness and structure specifically on this invisible null-space, leading to more accurate and less biased reconstructions.

The technical core of GSNR involves constructing a 'null-restricted Laplacian' from a graph representing pixel similarities and projecting the null-space signal onto its smoothest spectral modes. This approach offers three key advantages: improved algorithmic convergence, better coverage of null-space variance, and high predictability of the missing components from the measurements. The researchers demonstrated GSNR's effectiveness by integrating it into established solvers like Plug-and-Play (PnP) and Deep Image Prior (DIP) for tasks including image deblurring, compressed sensing, demosaicing, and super-resolution. The results are compelling, with GSNR providing consistent improvements of up to 4.3 dB in Peak Signal-to-Noise Ratio (PSNR) over baseline methods and up to 1 dB over some end-to-end learned models, marking a substantial advance in computational imaging.