A University of Manchester student has developed a powerful new ultra‑lightweight tool that can turn dark, noisy footage into clear, detailed and usable images.
Multinex, a new model for low‑light image enhancement (LLIE), was created by Computer Science undergraduate Alexandru Brateanu during his third-year project, working with academic supervisors.
The model outperforms comparable compact systems, recovering detail and clarity from images that would previously have been considered unusable.
The advancement has significant implications for photography, security, and a wide range of computational imaging tasks.
Low‑light image enhancement seeks to restore natural visibility, colour fidelity, and structural detail in scenes captured under poor illumination. While recent LLIE models have achieved impressive results, many rely on heavy architectures with large parameter counts, resulting in high computational cost and limited real‑time applicability. Efficiency has therefore become a central research challenge: how to enhance images more effectively while dramatically reducing model size.
In the work presented at the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026, the team proposes a structured solution grounded in classical colour vision theory and implemented using modern neural components within the Retinex framework. Retinex, a foundational approach in image enhancement, decomposes an image into illumination (light) and reflectance (colour) components to better handle low‑light scenes.
The design motivation behind Multinex is to extract as much useful information as possible from low‑light images using a highly compact architecture. By prioritising enhancement over reconstruction and leveraging lightweight neural operations, Multinex achieves strong illumination correction, detail recovery, and colour fidelity while using only a fraction of the parameters required by existing approaches.
The model is released in both a lightweight version (45K parameters) and an extremely compact nano version (0.7K parameters), each offering substantial reductions in computational load. Comparison to corresponding lightweight models such as PairLIE (330K parameters) and ZeroDCE (80K parameters) Multinex shows a significant performance improvement.
Like other LLIE techniques, Multinex still faces challenges in scenes with severe spectral distortions, lens flares, or mixed artificial and natural lighting. The team aims to extend the framework to these complex cases, exploring alternative formulations such as tone‑mapping or multiplicative residuals, and applying Multinex principles to related domains including intrinsic image decomposition, colour constancy, underwater enhancement, and haze removal.
The researchers demonstrate that Multinex delivers state‑of‑the‑art performance at real‑time cost, highlighting the power of combining analytic priors with modern lightweight design.
