HVR-SSLE

Published in IEEE Access 2026 — see the final paper version here: DOI: 10.1109/ACCESS.2026.3665009

Abstract

Low-light image enhancement (LLIE) is a fundamental problem in computational photography, aiming to restore images degraded by coupled noise, color distortion, and detail loss under poor illumination. While recent Transformer- and diffusion-based approaches can improve perceptual quality, they often incur substantial computational costs and may generalize poorly when trained on limited paired benchmarks. In this work, we cast LLIE as an iterative, stepwise refinement process and propose HVR-SSLE (Hierarchical Visual Reasoning for Self-Supervised Low-light image Enhancement), a compact recursive framework that alternates low-level local refinement and high-level global restoration modules for progressive enhancement. Despite having only 0.34M parameters, HVR-SSLE delivers progressive refinement without the heavy computational burden of existing models. We further quantify scene-level train–test overlap in widely used benchmarks such as LOL-v1/v2, highlighting potential evaluation bias. To reduce reliance on LLIE-specific paired data, we train HVR-SSLE in a self-supervised manner on the general-purpose COCO dataset. Diverse low-light inputs are synthesized by applying parametric gamma-based degradation curves to normal-light images. The resulting model is then evaluated in a zero-shot setting on standard LLIE benchmarks. Trained solely on COCO, HVR-SSLE achieves competitive PSNR/SSIM on paired benchmarks such as LOL-v2 (Synthetic) and LSRW, and yields favorable no-reference scores (e.g., NIQE and PIQE) on real-world unpaired datasets including DICM and LIME, indicating strong cross-dataset generalization.

Architecture

Requirements

• CUDA 12.x + PyTorch 2.4.0 (GPU recommended).
• Install deps:

  pip install --no-cache-dir -r requirements.txt
  pip install --no-cache-dir opencv-python-headless==4.8.0.74

• Disable wandb if desired: export WANDB_DISABLED=true.

Docker

• Build:

  docker build -t hvr-ssle:latest -f dockerfile .

• Run:

  docker run --gpus all -v $PWD:/workspace/app -it hvr-ssle:latest

• Workdir in container: /workspace/app.

Data Layout

• COCO (for training): dataset/coco/train2017, dataset/coco/val2017, annotations at dataset/coco/annotations/instances_{train,val}2017.json.
• Eval sets (examples):
  • LOLv1: dataset/LOLv1/eval15/{low,high}
  • LOLv2 Real: dataset/LOL-v2/Real_captured/Test/{Low,Normal}
  • LOLv2 Synthetic: dataset/LOL-v2/Synthetic/Test/{Low,Normal}
  • LSRW: dataset/LSRW/Eval/{Huawei,Nikon}/{low,high}
• Adjust paths in exp_config/config.py (coco_*, eval_dataset).

Training

• Basic launch (2 GPUs example):

  accelerate launch --num_processes 2 --gpu_ids 0,1 \
    train.py \
    --config_path exp_config/config.py \
    --config_name Config \
    --exp_name short_test

• Resume:

  accelerate launch ... train.py \
    --config_path exp_config/config.py \
    --config_name Config \
    --exp_name short_test \
    --resume_from_checkpoint checkpoint/HVR-SSLE_short_test_xxxxx/checkpoint_epoch_latest

• Turn off wandb logging: prefix command with WANDB_DISABLED=true (or set in .env).

Inference

• Single image or folder:

  python infer_image.py \
    --config_path checkpoint/config.json \
    --weights_path checkpoint/HVR.safetensors \
    --input_path path/to/img_or_dir \
    --output_dir results/out \
    --device auto

• Notes:
  • Uses torch.compile when available (PyTorch 2.x); falls back to eager mode if not.
  • Runs forward repeatedly for N_supervision steps (from config + overrides) instead of model.sample.
  • Autocast is enabled on CUDA when autocast_dtype is bf16 or fp16 (bf16 requires hardware support).
• Options: --check_gpu_mem (dummy run to print GPU memory), overrides for iterative steps (--hvr_t, --hvr_c, --hvr_n_supervision, --hvr_n_sup_factor).
• Example script: edit paths in infer_image.sh and run bash infer_image.sh.

Outputs

• Training: checkpoint/<run>/ holds config.json, checkpoint_epoch_latest, best_model, wandb_id.txt.
• Inference: saved to --output_dir with original filenames.

Analysis

• SSIM-based train/test overlap analysis is in analysis/ (CSV matrices, derived match lists, and a heatmap PDF). See analysis/README.md for details.

Troubleshooting

• wandb 401: set WANDB_DISABLED=true or run wandb login with valid token.
• Port conflicts: change --main_process_port in accelerate launch.
• Data path errors: verify paths in exp_config/config.py.

Citation

If you find this work useful, please cite:

@ARTICLE{11396663,
  author={Choo, Dongwon and Deng, Qikang and Park, Taewon and Lee, Dohoon},
  journal={IEEE Access}, 
  title={HVR-SSLE: Hierarchical Visual Reasoning for Self-Supervised Low-Light Image Enhancement}, 
  year={2026},
  volume={14},
  number={},
  pages={34705-34725},
  keywords={Cognition;Visualization;Lighting;Image color analysis;Image enhancement;Computational modeling;Image restoration;Transformers;Benchmark testing;Adaptation models;Low-light image enhancement (LLIE);self-supervised learning;zero-shot generalization;hierarchical visual reasoning;Gamma correction},
  doi={10.1109/ACCESS.2026.3665009}
}