Try LumiNet
Experience low-light enhancement and denoising with a single-stage compact CNN.
Resizes image to 1024px max side (faster inference)
Drop Image Here
- or -
Click to UploadQualitative Results
Sample visual comparisons demonstrating joint enhancement and denoising on unseen images.
Quantitative Results
| Dataset | Task | PSNR ↑ | SSIM ↑ | LPIPS ↓ | Exposure Error ↓ | Colour Error ↓ | Edge Detail ↑ |
|---|---|---|---|---|---|---|---|
| LOLv2-Real | Enhancement | 21.13 | 0.84 | 0.18 | 0.07 | 0.01 | 0.81 |
| LSRW | Enhancement | 17.67 | 0.56 | 0.27 | 0.10 | 0.01 | 0.67 |
| DLL | Enhancement | 18.88 | 0.57 | 0.33 | 0.09 | 0.01 | 0.70 |
| SIDD | Denoising | 35.23 | 0.91 | 0.18 | - | 0.00 | 0.79 |
| NLL | Joint | 17.97 | 0.56 | 0.37 | 0.10 | 0.01 | 0.66 |
LumiNet Architecture
A lightweight single-stage CNN designed for joint low-light image enhancement and denoising.
TM-JRB (Tri-Map Joint Restoration Block)
Learns three degradation-aware guidance maps in the bottleneck:
- L - illumination map
- sigma - noise map
- E - structure map
Provides:
- brightness enhancement
- noise suppression
- structure preservation
SNR-Gated Skip Fusion (SGSF)
Controls feature transfer between encoder and decoder stages.
- reduces noisy feature propagation
- preserves reliable structural information
- improves reconstruction quality
Frequency-Aware Refinement
Combines spatial and frequency-domain restoration.
- Wavelet-inspired brightness refinement enhances low-frequency illumination information and preserves image details
- FFT amplitude refinement refines frequency amplitudes, preserves phase information, and maintains structural consistency
Tri-Condition Spatial Modulation (TCSM)
Uses L, sigma, and E to adapt feature restoration across different image regions.
- enhances dark regions
- suppresses noisy regions
- preserves edges and textures
About LumiNet
LumiNet is a lightweight single-stage image restoration framework for low-light image enhancement and denoising. The framework improves image visibility and visual quality under difficult lighting conditions while keeping computational cost low. The compact architecture supports efficient inference across enhancement, denoising, and joint restoration tasks.
The Problem
Low-light images commonly suffer from:
- poor visibility
- noise contamination
- colour distortion
- loss of fine details
Many existing methods also face practical limits:
- multi-stage processing pipelines
- high computational complexity
- increased inference time
- limited suitability for edge deployment
LumiNet Solution
LumiNet addresses these challenges through:
- joint enhancement and denoising within a single forward pass
- adaptive restoration guided by illumination, noise, and structure information
- frequency-aware refinement for brightness correction and detail preservation
- efficient architecture suitable for practical deployment