Gradient-Aware Gamma Adjustment for Robust Diffusion-Based Computational Imaging

Restoring images degraded by spatially variant distortions remains a fundamental challenge in computational imaging. While diffusion-based methods like Denoising Diffusion with Preconditioned Guidance (DDPG) offer efficient solvers, they typically rely on a static, globally fixed trade-off between Back-Projection (BP) and Least-Squares (LS) guidance. This static scheduling prevents the sampling process from adapting to heterogeneous local signal-to-noise characteristics, often leading to suboptimal restoration in complex or mixed-content regions. We propose Gradient-Aware Gamma Adjustment (GAGA), a spatially adaptive framework that modulates the preconditioned guidance in a pixel-wise manner. By leveraging multi-scale gradient cues extracted from intermediate reconstructions, GAGA dynamically adjusts the BP–LS balance to better accommodate local structural variations. From a theoretical perspective, we refine the LS guidance by explicitly aligning it with the Gaussian likelihood score via the noise-dependent scaling factor$\sigma _{e}^{-2}$. We further show that the resulting spatially adaptive preconditioner remains strictly positive definite under bounded modulation, guaranteeing well-posed and numerically stable guidance for ill-posed inverse problems. Extensive experiments on CelebA-HQ and ImageNet, as well as on spatially variant degradations and computational imaging tasks including super-resolution, inpainting, CT, and MRI reconstruction, demonstrate consistent improvements in both fidelity (PSNR) and perceptual quality (LPIPS) over static preconditioned baselines, without introducing noticeable computational overhead.