Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement (ICCV 2025)

Official PyTorch implementation of the paper "Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement", accepted by ICCV 2025 (High light).

🌟 Introduction

Light field (LF) imaging in low-light environments suffers from severe noise and detail loss. While standard enhancement methods focus on individual views, they often break the View Consistency—the fundamental geometric relationship between sub-aperture images (SAIs).

Our method introduces a Scene-Adaptive framework that restores high-quality LF images while strictly preserving their 4D geometric structure, even in extreme darkness.

Key Contributions:

1. Scene-Adaptive Luminance Mapping: Dynamically adjusts enhancement parameters based on global and local illumination statistics.
2. Spatial-Angular Interaction (SAI) Module: Deeply integrates cross-view information to compensate for noise-corrupted pixels.
3. EPI-based Geometric Constraint: A novel loss function targeting Epipolar Plane Images (EPI) to ensure smooth disparity and flicker-free view transitions.

---

📦 Dataset

The dataset used in this project is built upon the work from the following paper. If you use this dataset, please also cite their work:

Harnessing multi-view perspective of light fields for low-light imaging [l3fnet]

You can download our organized dataset from the following link:

• Baidu Netdisk (百度网盘): [L3F-data.zip]
• Extraction Code (提取码): iaec

Please add the following citation when using the dataset

  @ARTICLE{l3fnet,
  author={M. {Lamba} and K. K. {Rachavarapu} and K. {Mitra}},
  journal={IEEE Transactions on Image Processing}, 
  title={Harnessing Multi-View Perspective of Light Fields for Low-Light Imaging}, 
  year={2021},
  volume={30},
  pages={1501-1513},
  doi={10.1109/TIP.2020.3045617}}

---

🛠️ Environment

This code was implemented and tested with Python and PyTorch. Based on the provided requirements, the main environment dependencies are:

• Python: >= 3.7 (Recommended 3.8 based on torch version)
• PyTorch: == 1.10.1
• Torchvision: == 0.11.2

To install all required dependencies, please run:

pip install -r requirements.txt

---

🚀 Training

To train the model from scratch, you can use the following command:

python train.py --data_dir path/to/your/lowlightdata --output_dir path/to/save/outs --dataset all --epochs 10000 --learning_rate 1e-4 --gpu_no 0 --ckpt "" --patch 128 --n_view 5

Training Arguments:

• --data_dir : Directory path of the training dataset (default is ).
• --output_dir : Output path for saving training results and model weights (default is ).
• --dataset : Specifies the dataset used for training (default is all, meaning all data is used). It can be changed to 1_20, 1_50, or 1_100.
• --epochs : Total number of training epochs (default is 10000).
• --learning_rate : Initial learning rate of the model (default is 1e-4).
• --gpu_no : Specifies the GPU ID to use (default is "0").
• --ckpt : Path to the checkpoint weight file for a pre-trained model or resuming training (default is empty "").
• --patch : Patch size for random image cropping during training (default is 128).
• --n_view : Number of input views for the light field images (default is 5).

🧪 Testing

To evaluate the trained model, run the following command:

python test.py --data_dir path/to/your/LowLightLF/ --output_dir path/to/save/outs --gpu_no 0 --patch 128 --n_view 5

测试参数说明 (Testing Arguments):

Testing Arguments:

• --data_dir : Directory path of the testing dataset (default is ).
• --output_dir : Output path for saving testing results, such as the enhanced images (default is ).
• --gpu_no : Specifies the GPU ID to use during testing (default is 0).
• --patch : Patch size for image processing during testing (default is 128).
• --n_view : Number of input views for the light field images during testing (default is 5).

📝 Citation

If you find our work useful, please cite:

@InProceedings{Zhang_2025_ICCV,
    author    = {Zhang, Shuo and Gao, Chen and Lin, youfang},
    title     = {Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2025}
}

---

📱 Contact

For any questions, please contact Shuo Zhang at [zhangshuo@bjtu.edu.cn]. More research from our lab can be found at [INSIS-CV Page] .

Scan the QR code to Learn more details