elksun98/gicd-net

GICD-Net is a research project, written in PyTorch, that trains a neural network to make low light photos and hazy photos look brighter and clearer. The README, which is written in Chinese, says the codebase contains three things stacked together: a basic enhancement network, an extra branch called IPB that uses prior knowledge about lighting, and a third branch based on a conditional diffusion model that adds residual detail. There are also scripts for training, running the model on new images, and measuring image quality. The layout of the project follows a typical PyTorch research repository. lowlight_train.py is the training entry, lowlight_test.py runs inference and computes PSNR, SSIM, and LPIPS scores against ground truth. Other files cover the dataset loader, the model itself, custom loss functions, a high frequency and low frequency multi scale constraint module, no reference image quality scripts that compute TOPIQ, MANIQA, LIQE, MUSIQ, TReS, BRISQUE, and Contrast, an ablation study script, and a separate diffusion subfolder. Local data and generated outputs are excluded from git. Setup needs Python 3.8 or higher and a CUDA build of PyTorch. The README walks through making a virtual environment with python -m venv and installing requirements.txt, with separate notes for Linux, macOS, and Windows PowerShell. The training data has to be in paired folders: a low light or hazy input image and a matching ground truth image with the exact same filename. Two datasets are mentioned, LOL for low light and NHC for haze, each laid out under data with train and test subfolders. The README provides many example command lines. There are commands for training the base model on LOL, training with IPB enabled, joint end to end training of all three components, and training the diffusion model on top of an existing backbone. The inference section shows how to run with paired ground truth, how to run on unlabeled images, how to save intermediate outputs, and how to use only the backbone without the diffusion branch. The same patterns repeat for the NHC haze dataset.