HVI-SMM: Spatially-Adaptive Correction for Low-Light Image Enhancement

HVI-SMM establishes a new Real Time SOTA model, outperforming HVI-CIDNet(CVPR2025) model.

Overview

HVI-SMM is a novel low-light image enhancement framework that introduces a pixel-wise spatially-adaptive correction method based on the HVI color space. Unlike previous global correction approaches, HVI-SMM dynamically generates saturation and brightness correction maps for each pixel, enabling sophisticated local restoration and robust performance across diverse illumination conditions.

Key Features

• Spatial Modulation Module (SMM): Predicts pixel-wise correction intensities for both saturation and brightness, allowing fine-grained local enhancement.
• HVI Color Space: Utilizes the HVI color space to overcome the limitations of RGB/HSV in low-light scenarios, reducing color distortion and noise.
• Intensity Mean Loss: Introduces a loss function that adaptively balances brightness alignment and detail restoration, focusing the network on structural fidelity.
• Intermediate Supervision: Employs intermediate loss to resolve scale ambiguity and stabilize training.

Project Structure

• scripts/ : Training, evaluation, and utility scripts
  • train.py : Main training pipeline (multi-dataset, distributed, SMM/HVI options)
  • eval.py : Evaluation and metric computation
  • options.py : Argument parser for all training/eval options
  • utils.py : Logging, checkpoint, and helper functions
• net/ : Model architectures (CIDNet, CIDNet_SSM, HVI transforms, LCA, etc.)
• loss/ : Loss functions (including Intensity Mean Loss, Perceptual, Edge, etc.)
• data/ : Dataset loaders and augmentation
• results/ : Output images, logs, and evaluation results
• weights/ : Pretrained and training checkpoints

Training

Example command (see scripts/options.py for all options):

python scripts/train.py --dataset lolv2_syn --batchSize 4 --nEpochs 500 --model_file net/CIDNet_SSM.py --use_gt_mean_loss hvi --use_random_gamma True

• Supported datasets: LOL-v1, LOL-v2 (real/syn), SID, SICE, LOL-Blur, etc.
• Multi-GPU: Automatically uses all available GPUs

Evaluation

• Run scripts/eval.py to evaluate trained models on supported datasets.
• Metrics: PSNR, SSIM, LPIPS, and custom metrics (see scripts/measure.py).

Citation

If you use HVI-SMM in your research, please cite:

@article{lee2025hvismm,
  title={HVI-SMM: Spatially-Adaptive Correction for Low-Light Image Enhancement},
  author={Injae Lee and Sungho Kang and Juneho Yi},
  journal={TODO},
  year={2025}
}

Acknowledgements

• Based on the HVI-CIDNet framework (original repo).
• Thanks to the authors of HVI-CIDNet and the open-source community.