L²DM: A Diffusion Model for Low-Light Image Enhancement

This paper presents L²DM, a novel framework for low-light image enhancement using diffusion models.

Requirements

A suitable conda environment can be created and activated with:

conda env create -f environment.yaml
conda activate ldm

Data preparation

We used the LOL and LOL-v2 datasets, where the LOL-v2 dataset is divided into two parts: real and synthetic. The dataset and model weights are placed in Baidu Cloud for downloading. Dataset files should be placed inside the data\

Pretrained Models

We need 3 network checkpoints, which are Auto-encoder checkpoints, COCO pre-training weights, and dataset-correlated weights. Once downloaded, put the model.ckpt to ckpt/vq-f4/model.ckpt; epoch=000099.ckpt to ckpt/coco/epoch=000099.ckpt; <lol_>.ckpt to ckpt/<lol_>.ckpt

Training L²DM

In configs/latent-diffusion/ we provide configs for training L²DM on the LOL, LOL-real, LOL-synthetic datasets. Training can be started by running CUDA_VISIBLE_DEVICES=<GPU_ID> python main_ll.py --base configs/latent-diffusion/<config_spec>.yaml -t --gpus 0,

Testing our results

Runpython d2l_ori.py --id 0 --dataset v1 --steps 20 --nrun 10 --sample dpm ,
dataset are available in v1, v2-real, and v2-syn.