AdaQual-Diff: Diffusion-Based Image Restoration via Adaptive Quality Prompting

TL;DR

AdaQual-Diff tackles the challenge of restoring images with spatially non-uniform real-world degradations by integrating perceptual quality scores into diffusion-based restoration. It introduces an Adaptive Quality Prompting mechanism that maps regional quality scores $Q$ from DeQAScore to prompt complexity $C_p$, creating a global prompt $P_g$ and region-specific prompts $P_l(Q)$ to form a spatially varying guidance field. The method formalizes the relationship $C_p \propto f(1-Q)$, employs a quality-weighted loss $L_{total} = L_{noise} + \lambda_1 L_{quality} + \lambda_2 L_{percep}$ with spatial weights $w(q)$, and uses a cache for DeQAScore to maintain practicality. Empirically, AdaQual-Diff achieves state-of-the-art results on composite degradation (CDD-11) and adverse-weather benchmarks, while maintaining efficiency with only 2 diffusion steps and a compact parameter count of 61.12M, demonstrating practical impact for real-world restoration tasks and downstream perception systems.

Abstract

Restoring images afflicted by complex real-world degradations remains challenging, as conventional methods often fail to adapt to the unique mixture and severity of artifacts present. This stems from a reliance on indirect cues which poorly capture the true perceptual quality deficit. To address this fundamental limitation, we introduce AdaQual-Diff, a diffusion-based framework that integrates perceptual quality assessment directly into the generative restoration process. Our approach establishes a mathematical relationship between regional quality scores from DeQAScore and optimal guidance complexity, implemented through an Adaptive Quality Prompting mechanism. This mechanism systematically modulates prompt structure according to measured degradation severity: regions with lower perceptual quality receive computationally intensive, structurally complex prompts with precise restoration directives, while higher quality regions receive minimal prompts focused on preservation rather than intervention. The technical core of our method lies in the dynamic allocation of computational resources proportional to degradation severity, creating a spatially-varying guidance field that directs the diffusion process with mathematical precision. By combining this quality-guided approach with content-specific conditioning, our framework achieves fine-grained control over regional restoration intensity without requiring additional parameters or inference iterations. Experimental results demonstrate that AdaQual-Diff achieves visually superior restorations across diverse synthetic and real-world datasets.

Figures (12)

• Figure 1: Our analysis reveals a mathematical relationship between perceptual quality metrics and optimal prompt complexity. We formalize this as $C_p \propto f(1-Q)$, where $C_p$ represents prompt complexity (token count) and $Q$ denotes regional quality score. AdaQual-Diff implements this relationship by decomposing the guidance into a global quality prompt $P_g$ and a local repair prompt $P_l(Q)$ whose complexity scales inversely with quality. This approach enables precision-targeted allocation of computational resources during the diffusion process.
• Figure 2: Overview of our AdaQual-Diff framework. Left: The complete pipeline incorporates region-wise quality assessment using DeQAScore, which evaluates degraded inputs on a scale of $q \in [1-5]$. The resulting quality map guides our Adaptive Quality Prompting mechanism, which dynamically generates text prompts whose complexity correlates with degradation severity. Right: Illustration of our quality-prompt relationship, where prompt complexity (token length) is inversely proportional to image quality—detailed, extensive prompts for severely degraded regions (low DeQAScore) and concise prompts for high-quality regions.
• Figure 3: Quality assessment visualization. (a) An example degraded image, (b) 3D representation of the quality map showing spatial distribution of quality scores across the image where brighter regions (yellow) indicate higher quality areas and darker regions (blue) represent lower quality areas affected by degradation. The quality scores range from 1 (lowest) to 5 (highest).
• Figure 4: Comparison of image restoration on low+haze+rain (the upper two examples) and low+haze+snow (the lower two examples) samples in real-world scenarios.
• Figure 5: Comparison of image restoration under real-world adverse weather conditions.