New SNORE method improves image restoration by matching noise levels
Denoisers now get the right noise level, outperforming diffusion models on deblurring.
A team from CNRS and Université Côte d'Azur (Marien Renaud, Jean Prost, Arthur Leclaire, Nicolas Papadakis) has proposed SNORE (Stochastic deNOising REgularization), a new algorithm that tackles image restoration by fixing a fundamental inconsistency in plug-and-play (PnP) methods. Traditional PnP methods iteratively apply a denoising neural network to images that become less noisy each step—a mismatch from the denoiser's training conditions. In contrast, diffusion models (DMs) denoise only after adding controlled noise, but at higher compute cost.
SNORE bridges the gap by using stochastic regularization: at each iteration, a carefully scaled noise is added to the current estimate before passing it through the denoiser. This ensures the denoiser always sees noise at the level it was trained on. The algorithm is formulated as a stochastic gradient descent (SGD) on an explicit objective, enabling convergence guarantees. An annealing version gradually reduces the noise schedule for finer results.
Experiments on standard deblurring and inpainting benchmarks show SNORE achieves competitive or superior PSNR and SSIM metrics compared to leading diffusion-based and PnP methods, with faster runtime and simpler hyperparameter tuning. The authors provide convergence proofs and open-source code.
- SNORE applies denoisers only at their trained noise level using explicit stochastic regularization, fixing a key flaw in prior PnP methods.
- The algorithm is proven to converge and includes an annealing schedule for improved restoration quality.
- Matches or exceeds state-of-the-art diffusion-based methods on deblurring and inpainting, with simpler implementation.
Why It Matters
A cleaner, theoretically grounded approach to image restoration that could replace heavier diffusion models in real-world applications.