HIT-LLIE-team solution for the NTIRE 2026 Efficient Low-Light Image Enhancement Challenge
Challenge Report | 中文文档 | Quick Start | Results | Reproducibility | Citation
MobileIE-6Ch is an ultra-lightweight low-light image enhancement network designed for efficient deployment. The model is based on the MobileIE family and introduces a 6-channel Retinex-style prediction head that estimates:
- 3 RGB illumination channels for color-aware brightness recovery
- 3 residual channels for noise/detail compensation
The released checkpoint contains only 101,922 parameters, while still targeting visually pleasing enhancement under strict NTIRE efficiency constraints. The repository includes inference code, DDP training code, configuration files, and the submitted checkpoint.
Suggested GitHub repository description
Official PyTorch implementation of MobileIE-6Ch, HIT-LLIE-team solution for NTIRE 2026 Efficient Low-Light Image Enhancement. 101.9K-parameter Retinex-style model with pretrained checkpoint.
The examples below show low-light inputs and the corresponding MobileIE-6Ch enhanced outputs. The model improves scene brightness while recovering visible structure in indoor, outdoor, street, and object-detail scenes.
| Design Goal | What We Do |
|---|---|
| Compact inference | Use the slim reparameterized MobileIE form for deployment |
| Color-aware illumination | Predict RGB illumination instead of a single grayscale map |
| Noise/detail recovery | Add an RGB residual branch after Retinex division |
| Training capacity | Use Multi-Branch Reparameterization during training |
| Stable optimization | Combine warm-up, EMA, gradient clipping, cosine scheduling, and multi-scale crops |
MobileIE-6Ch is listed in the official NTIRE 2026 Efficient Low-Light Image Enhancement Challenge report.
| Evaluation Table | SSIM | LPIPS | DISTS | LIQE | MUSIQ | Q-Align | Params | Final Rank |
|---|---|---|---|---|---|---|---|---|
| Main technical-report table | 0.5766 | 0.5176 | 0.2319 | 2.2977 | 60.3387 | 3.2007 | 101,922 | 7 |
| Full final-testing table | 0.5766 | 0.5176 | 0.2319 | 2.2977 | 60.3387 | 3.2007 | 101,922 | 9 |
The full final-testing table includes teams that participated in final testing but did not submit a technical report. The main table reports teams included in the technical-methods section of the challenge report.
MobileIE-6Ch follows a compact Retinex-inspired enhancement pipeline. Instead of estimating only one illumination channel, the model predicts RGB illumination maps and residual corrections jointly.
flowchart LR
A["Low-light RGB image"] --> B["MobileIE-6Ch backbone"]
B --> C["Dual attention modulation"]
C --> D["6-channel prediction"]
D --> E["RGB illumination map"]
D --> F["RGB residual map"]
A --> G["Retinex reconstruction"]
E --> G
F --> G
G --> H["Enhanced RGB image"]
The final image is reconstructed as:
enhanced = input / illumination + residual
During training, multi-branch convolution blocks improve representation capacity. For inference, the model uses the slim reparameterized form, keeping the checkpoint small and easy to deploy.
conda create -n mobileie python=3.10 -y
conda activate mobileie
pip install -r requirements.txtIf your CUDA version requires a specific PyTorch wheel, install the matching PyTorch build first, then install the remaining dependencies.
Put low-light images into:
competition/low/
python infer_6channel.pyEnhanced images will be saved to:
competition/enhanced_pt/
By default, the script loads the released checkpoint:
result/model_best.pt
The script uses CUDA automatically when available and falls back to CPU otherwise.
Prepare paired training data with matching filenames:
lowlight/
|-- low/ # Low-light inputs
`-- normal/ # Normal-light ground truth
Launch DDP training:
bash train_ddp.shChoose a custom GPU list if needed:
bash train_ddp.sh "0" 1
bash train_ddp.sh "0,1" 2
bash train_ddp.sh "0,1,2,3" 4The main configuration is in config/lle.yaml.
| Setting | Value |
|---|---|
| Model | MobileIE-6Ch |
| Channels | 32 |
| Epochs | 800 |
| Batch size | 16 |
| Warm-up | 20 epochs |
| Learning rate | 1.5e-4 |
| Scheduler | Cosine annealing |
| Patch size | 768 |
| Cross validation | 5 folds |
| EMA | Enabled |
| Gradient clipping | 0.5 |
Training logs and checkpoints are written under experiments/.
This repository includes the core materials needed to reproduce and inspect the submitted solution.
| Asset | Status | Notes |
|---|---|---|
| Source code | Included | Model, data loader, losses, metrics, inference, and DDP training |
| Pretrained checkpoint | Included | result/model_best.pt, 101,922 parameters |
| Training config | Included | config/lle.yaml |
| Dataset | Not redistributed | Please follow the NTIRE challenge dataset policy |
| Generated predictions | Not tracked | Place inputs in competition/low/ and outputs are saved to competition/enhanced_pt/ |
| Citation metadata | Included | CITATION.cff |
| Model card | Included | docs/MODEL_CARD.md |
| Reproduction notes | Included | docs/REPRODUCIBILITY.md |
For more details, see Reproducibility Notes and Model Card.
.
|-- .github/
| `-- ISSUE_TEMPLATE/ # Bug report and question templates
|-- docs/
| |-- MODEL_CARD.md # Model usage, limitations, and intended scope
| `-- REPRODUCIBILITY.md # Reproduction checklist and environment notes
|-- competition/
| |-- low/ # Inference inputs
| `-- enhanced_pt/ # Inference outputs
|-- config/
| `-- lle.yaml # Training and model configuration
|-- data/
| |-- lledata.py # Low-light enhancement dataset loader
| `-- ispdata.py
|-- lowlight/
| |-- low/ # Training low-light images
| `-- normal/ # Training ground-truth images
|-- model/
| |-- lle.py # Original MobileIE LLE model
| |-- lle_6channel.py # MobileIE-6Ch model
| `-- utils_IWO.py # MBR and feature modulation blocks
|-- result/
| `-- model_best.pt # Released MobileIE-6Ch checkpoint
|-- infer_6channel.py # Inference entry point
|-- main_ddp.py # DDP training entry point
|-- train_ddp.sh # Training launcher
|-- CITATION.cff # Machine-readable citation metadata
|-- requirements.txt
`-- team_info.txt
HIT-LLIE-team
| Member | GitHub | Affiliation |
|---|---|---|
| Xinbai Wang | w-xb | Harbin Institute of Technology |
| Duo Liu | Cat-blizzard | Harbin Institute of Technology |
Contact information is available in team_info.txt.
If this repository is helpful to your research, please cite the NTIRE 2026 challenge report and this solution.
@article{yan2026ntire,
title={NTIRE 2026 Challenge on Efficient Low Light Image Enhancement: Methods and Results},
author={Yan, Jiebin and Tu, Chenyu and Lin, Qinghua and others},
journal={arXiv preprint arXiv:2605.02212},
year={2026}
}@misc{hitllie2026mobileie6ch,
title={MobileIE-6Ch: Efficient Low-Light Image Enhancement for NTIRE 2026},
author={Wang, Xinbai and Liu, Duo},
year={2026},
note={HIT-LLIE-team submission to the NTIRE 2026 Efficient Low-Light Image Enhancement Challenge}
}This project was developed for the NTIRE 2026 Efficient Low-Light Image Enhancement Challenge. We thank the challenge organizers and the MobileIE authors for the baseline inspiration.
This repository is released under the Apache License 2.0. See LICENSE for details.