SMALL: Baseline Models and Checkpoints

Code and 12 best-validation checkpoints for Retinexformer, CPGA-Stable, CIDNet and FusionNet, with training crops 256, 512 and 1024.

source/: architectures and inference code.
weights/: best checkpoints; model/crop/epoch/hash are listed in checkpoints.csv.
config.json: settings declared in manuscript Section 5.2, provided as documentation rather than original run logs. This package provides inference code, not a training pipeline.

Usage

pip install -r requirements.txt
python source/infer.py --model Retinexformer --checkpoint weights/Retinexformer/crop1024/best.pt --input input.png --output enhanced.png --device cuda

Use --device cpu for CPU inference. Supported model names: Retinexformer, CPGA-Stable, CIDNet, FusionNet. Select the corresponding checkpoint for each model and crop. Images retain their original dimensions, with temporary padding removed after inference (multiples of 32 for FusionNet and 8 for the other models). When resizing or cropping paired images, apply identical resize dimensions and crop coordinates to both the low-light and reference images. Preserve the aspect ratio when resizing.

All 12 checkpoints passed strict loading locally. This check does not establish reproduction of every benchmark result. Original checkpoint files are preserved.

Dataset: https://data.mendeley.com/datasets/vnz4gcw8wk/1

Third-party license notices are retained in third_party/ where available. This repository does not replace the original authors' license terms.

Downloads last month: 34