| --- |
| language: |
| - en |
| library_name: pytorch |
| pipeline_tag: image-to-image |
| tags: |
| - low-light-image-enhancement |
| - low-light-enhancement |
| - llie |
| - image-enhancement |
| - image-restoration |
| - computer-vision |
| - pytorch |
| - retinex |
| - lightweight-model |
| - edge-ai |
| - real-time |
| - image-to-image |
| - exposure-correction |
| - illumination-enhancement |
| - color-correction |
| pretty_name: Multinex |
| thumbnail: assets/teaser.png |
| model-index: |
| - name: Multinex |
| results: |
| - task: |
| type: image-enhancement |
| name: Low-Light Image Enhancement |
| dataset: |
| name: LOL-v1 |
| type: lol-v1 |
| metrics: |
| - type: psnr |
| name: PSNR |
| value: 23.19 |
| - type: ssim |
| name: SSIM |
| value: 0.843 |
| - type: lpips |
| name: LPIPS |
| value: 0.129 |
| --- |
| |
| # Multinex: Lightweight Low-Light Image Enhancement via Multi-prior Retinex |
|
|
| **Multinex** is a lightweight model for **low-light image enhancement**. It is designed to brighten dark images, recover clearer colors, and improve visible detail while staying very small and efficient. |
|
|
| Multinex is built around a simple idea: instead of relying only on raw RGB input, it uses multiple useful image cues related to brightness and color, then combines them through a compact network. This helps it perform strongly even at a very small model size. |
|
|
| ## Model Description |
|
|
| Many low-light enhancement methods are either large and expensive to run, or they struggle to separate brightness changes from color correction cleanly. |
|
|
| Multinex addresses this with a lightweight and structured design. It uses: |
| - a brightness-oriented branch |
| - a color-oriented branch |
| - compact fusion modules |
| - carefully chosen analytic image priors |
|
|
| These priors give the model a stronger starting point, so it does not need to learn everything from scratch. |
|
|
| ## Why Multinex? |
|
|
| Multinex is useful when you want: |
| - strong low-light enhancement |
| - a small and efficient model |
| - better brightness and color recovery |
| - a method suitable for edge or real-time use |
| - a practical preprocessing step for downstream vision tasks |
|
|
| ### Main advantages |
| - **Lightweight**: very small compared to many recent methods |
| - **Efficient**: suitable for resource-constrained settings |
| - **Structured**: uses separate cues for brightness and color |
| - **Practical**: useful both for visual enhancement and downstream tasks |
| - **Robust**: performs well across both reference and no-reference benchmarks |
|
|
| ## How It Works |
|
|
| Multinex takes a dark RGB image and improves it using two types of guidance: |
| - **luminance guidance**, which helps with brightness |
| - **reflectance guidance**, which helps with color and appearance |
|
|
| These two forms of guidance are processed by a lightweight network and combined into the final enhanced image. |
|
|
| In simple terms, Multinex tries to answer two questions: |
| - how should this image be brightened? |
| - how should its colors and details be corrected? |
|
|
| By handling these separately, the model can enhance low-light images more effectively. |
|
|
| ## Prior Stacks |
|
|
| A key part of Multinex is its use of **prior stacks**. |
|
|
| ### Luminance guidance stack |
| This stack provides different views of image brightness. It helps the model understand where the image is dark, how light is distributed, and how brightness should be adjusted. |
|
|
| ### Reflectance guidance stack |
| This stack provides different views of image color and chromatic structure. It helps the model better recover color, preserve regions, and reduce unwanted artifacts. |
|
|
| Together, these stacks give the model a richer and more stable representation of the image. |
|
|
| ## Intended Uses |
|
|
| ### Direct use |
| - Enhance dark photos |
| - Improve brightness and color visibility |
| - Recover clearer low-light images |
| - Preprocess images before other vision tasks |
|
|
| ### Downstream use |
| - Machine vision |
| - Low-light object detection |
| - Surveillance and nighttime imaging |
| - Embedded and edge systems |
| - Mobile photography enhancement |
| - Robotics in dim environments |
|
|
| ## Training and Evaluation Data |
|
|
| Multinex is evaluated on common low-light image enhancement benchmarks, including: |
| - **LOL-v1** |
| - **LOL-v2-real** |
| - **LOL-v2-syn** |
| - **MEF** |
| - **LIME** |
| - **DICM** |
| - **NPE** |
| - **ExDark** for downstream detection evaluation |
|
|
| Please refer to the paper and code for exact training details and evaluation settings. |
|
|
| ## Performance |
|
|
| The paper reports that Multinex: |
| - outperforms earlier lightweight and micro models in several settings |
| - achieves strong no-reference perceptual quality |
| - remains competitive with much larger methods |
| - improves downstream object detection performance in low-light scenes |
|
|
| ### Example paired restoration results |
|
|
| | Dataset | PSNR | SSIM | LPIPS | |
| |---|---:|---:|---:| |
| | LOL-v1 | 23.19 | 0.843 | 0.129 | |
| | LOL-v2-real | 23.04 | 0.860 | 0.178 | |
| | LOL-v2-syn | 25.04 | 0.930 | 0.068 | |
|
|
| ### Example no-reference results |
|
|
| | Metric | Mean | |
| |---|---:| |
| | NIQE ↓ | 3.64 | |
| | BRISQUE ↓ | 15.89 | |
|
|
| ### Efficiency |
|
|
| GFLOPs computed on 256x256x3 inputs. |
|
|
| | Variant | Parameters | GFLOPs | |
| |---|---:|---:| |
| | Multinex | 0.0447M | 2.50 | |
| | Multinex-Nano | 0.0007M | 0.04 | |
|
|
|
|
| ## Citation |
|
|
| If you use Multinex in your work, please cite the paper. |
|
|
| ```bibtex |
| @inproceedings{multinex2026, |
| title = {Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex}, |
| author = {Alexandru Brateanu and Tingting Mu and Codruta O. Ancuti and Cosmin Ancuti}, |
| booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, |
| year = {2026} |
| } |
