Update README.md

README.md

@@ -1,3 +1,184 @@
----
-license: cc-by-nc-sa-4.0
----
+---
+language:
+- en
+license: mit
+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 **low-light image enhancement (LLIE)** model based on a **multi-prior Retinex** formulation. It is designed to improve **brightness, color fidelity, and structural detail** in dark images while remaining extremely compact and efficient.
+
+Multinex combines:
+- a **Retinex-guided residual enhancement formulation**
+- **analytic luminance and reflectance prior stacks**
+- a **dual-branch lightweight fusion network**
+
+This design allows the model to achieve strong perceptual enhancement quality and competitive paired-image restoration performance at a tiny parameter budget, including a lightweight version around **45K parameters** and an even smaller nano variant around **0.7K parameters**. The paper reports strong results on both enhancement and downstream detection benchmarks.  
+
+**Keywords:** low-light image enhancement, LLIE, Retinex, image restoration, lightweight neural network, edge deployment, exposure correction, color correction, perceptual image enhancement, real-time enhancement.
+
+## Model Description
+
+Many low-light image enhancement methods either:
+- use large neural networks that are hard to deploy on-device, or
+- operate in a single color space where illumination and color remain partially entangled.
+
+Multinex addresses this with a structured and lightweight design. Instead of learning a full reconstruction in a standard Retinex manner, it predicts an **enhancement delta** added to the input image. The correction is factorized into:
+- a **luminance correction**
+- a **reflectance correction**
+
+The model also uses **analytic representation priors** to reduce the burden on the network. In particular:
+- the **luminance guidance stack** provides illumination-oriented cues
+- the **reflectance guidance stack** provides chromatic and color-structure cues
+
+This makes the model especially attractive for:
+- mobile or embedded deployment
+- lightweight computer vision pipelines
+- preprocessing for downstream tasks such as object detection in dark scenes
+
+## Why Multinex?
+
+Multinex is useful when you need a **small and efficient low-light enhancement model** without giving up physical interpretability.
+
+### Main advantages
+- **Lightweight**: designed for very low parameter counts
+- **Retinex-inspired**: uses luminance/reflectance structure as a prior
+- **Analytic priors**: leverages hand-crafted luminance and chrominance descriptors
+- **Strong perceptual quality**: especially strong on no-reference perceptual metrics
+- **Practical utility**: enhancement can improve downstream visual tasks in low light
+
+## Architecture Overview
+
+Multinex enhances an RGB input image through a residual Retinex-based formulation:
+
+\[
+\hat{\mathbf{I}} = \mathbf{I} + \Delta_{\mathbf{I}} = \mathbf{I} + \Delta_L \odot \Delta_R
+\]
+
+where:
+- \(\mathbf{I}\) is the low-light input image
+- \(\hat{\mathbf{I}}\) is the enhanced image
+- \(\Delta_L\) is the learned luminance correction
+- \(\Delta_R\) is the learned reflectance correction
+
+Instead of learning directly from raw RGB alone, Multinex constructs two analytic prior stacks:
+- **Luminance stack** for brightness-related structure
+- **Reflectance stack** for color and chromatic structure
+
+These stacks are processed by lightweight fusion modules to produce the correction terms.
+
+## Prior Stacks
+
+A key idea in Multinex is that **the network should not need to rediscover useful image priors from scratch**.
+
+### Luminance guidance stack
+The paper uses a luminance stack built from multiple analytic luminance descriptors, including variants such as:
+- Rec.709 luminance
+- max-based luminance
+- lightness-based luminance
+- L2-based luminance
+
+### Reflectance guidance stack
+The reflectance stack uses complementary chromatic descriptors, including:
+- \(C_b, C_r\)
+- normalized chromaticities \(r, g\)
+- saturation \(S\)
+
+The supplementary further motivates these choices with **Linear Reconstruction Analysis (LRA)** and descriptor-complementarity analysis.
+
+## Intended Uses
+
+### Direct use
+- Enhance dark photos
+- Improve brightness and color visibility in low-light scenes
+- Preprocess images before downstream computer vision tasks
+
+### Downstream use
+- Machine Vision
+- Low-light object detection
+- Surveillance and nighttime vision preprocessing
+- Embedded or edge imaging systems
+- Mobile photography enhancement
+- Robotics and autonomous perception in dim environments
+
+## Training Data
+
+The paper evaluates 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 check the paper and codebase for the exact training split, preprocessing, and evaluation protocol used for each released checkpoint.
+
+## Performance
+
+The paper reports that Multinex:
+- outperforms prior lightweight and micro models in several settings
+- achieves strong no-reference perceptual performance
+- remains competitive relative to much larger methods
+- improves downstream object detection performance in low-light conditions
+
+### Example reported paired restoration results
+For the lightweight Multinex model, the paper reports:
+
+| 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 reported no-reference results
+| Metric | Mean |
+|---|---:|
+| NIQE ↓ | 3.64 |
+| BRISQUE ↓ | 15.89 |
+
+### Efficiency
+| Variant | Parameters | GFLOPs |
+|---|---:|---:|
+| Multinex | 0.0447M | 2.50 |
+| Multinex-Nano | 0.0007M | very low |