Students at the University of Manchester have developed a powerful new ultra-light tool that can transform dark, noisy footage into clear, detailed and easy-to-use images.
Multinex is a new model for low-light image enhancement (LLIE) created by computer science undergraduate Alexandru Brateanu during his third-year project in collaboration with his academic advisor.
This model outperforms comparable compact systems and restores image detail and clarity that was previously considered unusable.
This advancement has significant implications for photography, security, and a wide range of computational imaging tasks.
Low-light image enhancement aims to restore the natural visibility, color fidelity, and structural details of scenes captured in poor lighting conditions. Although recent LLIE models have achieved impressive results, many rely on heavy architectures with large parameter counts, resulting in high computational costs and limited real-time applicability. Therefore, efficiency has become a central research topic. In other words, it’s a way to enhance images more effectively while significantly reducing model size.
In research presented at the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026, the team proposes a structured solution based on classical color vision theory and implemented using modern neural components within the Retinex framework. Retinex, a fundamental approach to image enhancement, decomposes images into illumination (light) and reflectance (color) components to better handle low-light scenes.
“My interest began during a research internship after my first year of college, where I became increasingly focused on making visual AI smaller and smarter. Multinex grew from the idea that better problem formulations lead to more efficient models. The use of the Retinex principle with multiple descriptions of light and color allows a compact network to focus limited capacity on the augmentation task itself, making it suitable for real-time AI in safety-critical vision systems.” Mr Brateanu, Principal Researcher and Student at the University of Manchester.
The design motivation behind Multinex is to extract as much useful information as possible from low-light images using a very compact architecture. By prioritizing enhancement over reconstruction and leveraging lightweight neural operations, Multinex achieves powerful lighting correction, detail recovery, and color fidelity while using only some of the parameters required by existing approaches.
The model is released in both a lightweight version (45K parameters) and a very compact nano version (0.7K parameters), each of which significantly reduces the computational load. When compared to its lightweight counterparts, Multinex, such as PairLIE (330K parameters) and ZeroDCE (80K parameters), we see a significant performance improvement.
Like other LLIE technologies, Multinex still faces challenges in scenes with severe spectral distortion, lens flare, or a mix of artificial and natural light. The team aims to extend the framework to these complex cases, explore alternative formulations such as tone mapping and multiplicative residuals, and apply Multinex principles to related areas such as intrinsic image decomposition, color constancy, underwater enhancement, and haze removal.
The researchers demonstrate that Multinex provides state-of-the-art performance at real-time costs, highlighting its ability to combine up-front analysis with modern lightweight design.
“Low-light image enhancement is essential to world modeling, which is the foundation of next-generation AI. It enables stable, predictive representations of real-world environments where standard visual assumptions don’t hold.More broadly, this study highlights the importance of integrating classical knowledge about light, color, and perception into modern AI systems—to augment, not replace.” “This will be important for the system to achieve true autonomous, real-world operation,” he said. Tingting Mu, Associate Professor of Machine Learning, University of Manchester.
Source: University of Manchester
