Overlapped Wavelet Diffusion for Low-Light Image Enhancement

A research team has proposed a new overlapped wavelet diffusion framework designed to improve low-light image enhancement by eliminating blocking artifacts and preserving fine detail, according to a preprint submitted on 9 Jun 2026 [1]. The framework, called OWDiff, targets two specific shortcomings in the existing diffusion-based method DiffLL. DiffLL employs a Haar Wavelet Transform, which can introduce visible blocking artifacts, and a High-Frequency Restoration Module that often results in blurred edges or over-smoothed textures [1]. To address these issues, the researchers introduced an Overlapped Wavelet Transform that incorporates correlations across neighboring regions, structurally preventing the formation of blocking artifacts [1]. A second component, a low-frequency-guided High-Frequency Enhance Block, was integrated to strengthen detail recovery and produce sharper edges [1]. Noise reduction in signal processing is a long-standing challenge, as algorithms designed to remove noise can also distort the underlying signal [3]. In digital images, this distortion can manifest as a loss of texture or the introduction of artificial patterns, problems the OWDiff framework is explicitly designed to mitigate [1]. The work was evaluated on the LOLv1 and LOLv2-real datasets, standard benchmarks for low-light image enhancement [1]. The authors report that OWDiff consistently outperformed existing methods both in qualitative visual assessments and quantitative metrics [1]. Compared directly to DiffLL, OWDiff achieved an average PSNR gain of 0.58 dB, a 1.64% relative improvement in SSIM, and a 5.9% relative reduction in LPIPS across both datasets [1]. The paper was posted on arXiv, an open-access repository for electronic preprints in fields including computer science and electrical engineering that has been operating since 1991 and now receives about 24,000 submissions per month [10].