Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement
Authors: Kai Shang, Mingwen Shao, Chao Wang, Yuanshuo Cheng, Shuigen Wang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on LOL and LOLv2 datasets demonstrate that our method achieves state-of-the-art performance for the low-light image enhancement task. |
| Researcher Affiliation | Collaboration | Kai Shang1,2, Mingwen Shao1*, Chao Wang3, Yuanshuo Cheng1, Shuigen Wang4 1School of Computer Science and Technology, China University of Petroleum (East China), China 2Shandong Institute of Petroleum and Chemical Technology, China 3Re LER, AAII, University of Technology Sydney, Australia 4Yantai IRay Technologies Lt. Co., China |
| Pseudocode | Yes | Algorithm 1: MDMS Diffusion Model Training; Algorithm 2: MDMS Diffusion Model Sampling |
| Open Source Code | Yes | Codes are available at https://github.com/Oliiveralien/MDMS. |
| Open Datasets | Yes | Dataset. The proposed diffusion model is trained on the LOL dataset (Wei et al. 2018), and evaluated on both LOL and LOLv2 dataset (Yang et al. 2021). |
| Dataset Splits | No | The paper specifies training and testing splits for the datasets (e.g., "The LOL dataset contains 500 paired images, with 485 for training and 15 for testing."), but it does not explicitly provide separate validation dataset splits. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types) used to run the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as libraries or programming languages used (e.g., Python version, PyTorch version, CUDA version). |
| Experiment Setup | Yes | Schedules. For our diffusion model, the time-step T is set to 1,000 for the training stage and the implicit sampling step S is set to 25. Furthermore, our model achieves promising results using alternative step (S = 20, 10, 5, 4). For the noise schedule, α is linearly decreased from 0.999 to 0.98. Training details. We conduct training using 64 × 64 patches. To correspond with multi-scale sampling and enhance training diversity, we randomly crop patches of size 256 × 256, 128 × 128, and 64 × 64, and resized them to 64 × 64. We use the Adam optimizer with an initial learning rate of 2e−5. In addition to the time step t and patch size, we also add parameters including the top-left and bottom-right coordinates of the cropped patches during the parameter embedding to guide the training and sampling process. |