UW-Madison's SGPnP Beats Diffusion Solvers on Severely Corrupted MRI Scans
SGPnP plugs pretrained score-based diffusion models directly into PnP solvers without reverse diffusion sampling. Injecting Gaussian noise into the denoiser input helps iterates escape strict saddle points in ill-posed problems.
Takeaways
- SGPnP plugs pretrained score-based diffusion models directly into PnP solvers without reverse diffusion sampling.
- Injecting Gaussian noise into the denoiser input helps iterates escape strict saddle points in ill-posed problems.
- Decoupling injected noise level from denoiser conditioning level accounts for data-consistency artifacts.
- Hits 25.21 PSNR on FFHQ box inpainting versus 18.50 for stochastic PnP baseline SNORE.
- On 4x fastMRI, SGPnP reaches 32.54 PSNR, beating DPS (29.60) and DiffPIR (29.81).
- Code available at github.com/uw-cig/SGPnP ; works with ADMM, PGM, and RED variants.
Plug-and-play (PnP) methods reconstruct images from incomplete or corrupted measurements by alternating between a data-consistency update and a learned denoiser. Data consistency enforces what the scanner or camera measured; the denoiser supplies an image prior, replacing the proximal operator that a conventional optimizer would use. This pattern supports MRI reconstruction, deblurring, super-resolution, and inpainting without requiring an explicit formula for the prior.
A new study from researchers at UW-Madison, Washington University in St. Louis, and Los Alamos National Laboratory finds that conventional PnP methods lose accuracy as measurements become sparse or corruption grows severe. Their Stochastic Generative Plug-and-Play method, or SGPnP, combines a pretrained score-based diffusion model with fresh Gaussian noise at each iteration. It competes with specialized diffusion solvers while preserving the existing PnP optimization loop. The authors have released the source code .
Hard corruptions expose PnP’s blind spot
Standard PnP denoisers are usually trained to remove low-level additive Gaussian noise. Intermediate iterates from difficult inverse problems look different: large masks create block-shaped gaps, undersampled MRI produces aliasing, and deblurring leaves structured residuals. This distribution shift gives the denoiser inputs that differ sharply from its training data.
Low-noise denoisers also lack the broad generative behavior needed to infer large missing regions. Re-noising methods such as SNORE reduce the distribution mismatch by adding noise before denoising, but their low-noise models provide limited gains on strongly underdetermined tasks. Score-based diffusion models cover a much wider range of corruption levels and learn image structure at multiple scales.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
