SADGE: Spatially-Adaptive Diffusion Guided by Estimated Degradation for Image Restoration

Diffusion restoration faces a mismatch with real-world degradation: corruption is spatially heterogeneous, but diffusion noise is typically global. A noise level large enough to recover severe damage over-perturbs reliable regions, risking hallucinated content, and lowering it under-restores. We resolve this trade-off with a spatially-adaptive mean-reverting SDE, replacing the scalar stationary variance with a learned per-pixel temperature field set by a predicted degradation map. The map is trained to match a residual-based severity target that combines local residual energy with a high-pass measure of spatial variation, and it modulates both the diffusion dynamics and conditions the denoiser, allocating stochasticity to the severity of local corruption. Introduced train-test coupling strategy further accounts for the discrepancy between analytically derived training targets and predicted maps at inference. As a result the method concentrates generative restoration on corrupted regions while suppressing unnecessary changes elsewhere, reducing hallucinations, and improving preservation of intact structure.