Self-supervised Dynamic Heterogeneous Degradation Modeling for Unified Zero-Shot Image Restoration
Abstract Overview
This paper presents UP-ZeroIR, a self-supervised zero-shot image restoration framework that aims to handle heterogeneous degradations within a single diffusion-based pipeline. The central idea is to reparameterize different degradations in latent space using a compact, physically coherent generalized Gaussian distribution, so that restoration can be guided by explicit degradation modeling rather than implicit feature prompts alone. The method combines degradation-aware posterior sampling, a physically coherent degradation model, a degradation alignment loss, and a dynamic quality-refinement strategy that adapts the diffusion trajectory during inference. Experiments cover low-light enhancement, dehazing, denoising, and mixed degradations, with ablations analyzing the role of each major component.
Novelty
The distinctive contribution is the explicit modeling of diverse image degradations as a unified low-dimensional physical distribution in latent space, rather than treating degradations as black-box conditions. The paper also introduces a quality-driven dynamic refinement mechanism that adaptively revises the diffusion trajectory during zero-shot inference.
Results
The reported experiments show that UP-ZeroIR achieves the strongest quantitative results among the compared methods on low-light enhancement, dehazing, and denoising benchmarks, and also performs best on two mixed-degradation settings. The paper reports PSNR gains over the second-best baseline of 0.95 dB and 0.98 dB on LOLv1/LOLv2, 1.03 dB on HSTS dehazing, and 0.85 dB on Kodak24 denoising. Ablation studies further indicate that the physical degradation model, degradation alignment loss, and dynamic quality refinement each contribute to the final performance.
Key Points
- UP-ZeroIR models heterogeneous degradations with a compact generalized Gaussian parameterization in latent diffusion space, providing an explicit physically grounded control signal for zero-shot restoration.
- Across low-light enhancement, dehazing, denoising, and mixed degradations, the method outperforms the compared zero-shot posterior-sampling baselines and achieves the best reported benchmark numbers in the paper.
- Ablation and robustness analyses indicate that physically coherent degradation modeling, distribution-level alignment, and dynamic quality refinement improve convergence stability and help the method cope with entangled degradations and added noise.