Multinex: An ultra lightweight AI model advancing low light image enhancement
A University of Manchester student has engineered a highly compact artificial intelligence tool capable of transforming dark, low-light footage into clear, high-detail imagery. Known as Multinex, the model significantly outperforms traditional lightweight systems by utilizing classical colour vision theory combined with modern neural architecture. According to the University of Manchester, this advancement allows for real-time image enhancement using as few as 700 parameters, a fraction of the computational load required by existing industry standards.
How Does Multinex Improve Low-Light Visibility?
Multinex functions by decomposing images into their base illumination and reflectance components, a process rooted in the Retinex framework. By focusing on these fundamental properties of light and colour, the model avoids the heavy computational requirements of standard deep learning networks. Lead researcher Alexandru Brateanu stated that by using classical principles, the network can dedicate its limited capacity strictly to the enhancement task, making it suitable for safety-critical systems that require immediate, real-time data processing.
Most existing low-light enhancement models, such as PairLIE, require upwards of 330,000 parameters to function. Multinex achieves comparable or superior results with its “nano” version using only 700 parameters.
Why Does Computational Efficiency Matter for AI?
Efficiency is the primary barrier to deploying AI in autonomous systems and security hardware. According to Dr. Tingting Mu, Associate Professor in Machine Learning at the University of Manchester, the ability to “perceive and reason in the dark” with minimal energy consumption is critical for the next generation of autonomous technology. High-parameter models often suffer from latency, which can be dangerous in environments like self-driving vehicles or real-time surveillance, where every millisecond of processing time impacts safety.
How Do Existing Models Compare to Multinex?
The performance gap between Multinex and previous models is significant when measuring parameter counts against image clarity. Data presented at the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 highlights the following efficiency comparisons:
- Multinex (Nano): 0.7K parameters
- ZeroDCE: 80K parameters
- PairLIE: 330K parameters
While models like PairLIE and ZeroDCE have been the standard for low-light image enhancement (LLIE), the research team notes that Multinex delivers state-of-the-art results while drastically reducing the power and hardware requirements, making it more viable for edge computing devices.
What Are the Next Steps for Low-Light AI Research?
Despite its efficiency, Multinex currently struggles with scenes involving mixed artificial lighting, lens flares, or severe spectral distortions. The research team intends to address these limitations by exploring alternative formulations such as tone-mapping and multiplicative residuals. Expanding the frameworkâs scope to include related domains like underwater enhancement, haze removal, and colour constancy remains a priority for future development, as these fields rely on similar principles of intrinsic image decomposition.
For those interested in the technical implementation of image restoration frameworks, you can explore the latest research papers and proceedings from the IEEE/CVF Conference on Computer Vision and Pattern Recognition to see how these architectures are evolving.
Frequently Asked Questions
What is the main advantage of the Multinex model?
The primary advantage is its extreme efficiency; it achieves high-fidelity image restoration using significantly fewer parameters than traditional models, allowing for real-time performance on low-power hardware.
Can this technology be used for security cameras?
Yes. Because the model is designed to be lightweight and efficient, it is well-suited for integration into safety-critical visual systems, including security and autonomous vehicle sensors.
What are the current limitations of the system?
Multinex faces challenges when processing scenes with complex lighting issues, such as severe lens flares or mixed natural and artificial light sources, which the team is currently working to resolve.
Have questions about the future of AI-driven visual processing? Leave a comment below or subscribe to our weekly tech briefing to stay updated on the latest breakthroughs in machine learning.