Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models
Authors: Yan Huang, Xiaoshan Liao, Jinxiu Liang, Yuhui Quan, Boxin Shi, Yong Xu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our framework outperforms existing zero-shot methods across diverse real-world scenarios. Experiments Experimental Settings Datasets We conduct experiments on two widely used datasets with paired low-light and normal-light images, LOLv1 (Wei et al. 2018) and LOL-v2 (Yang et al. 2020). We also provide visual comparisons on datasets without ground truth (DICM (Lee, Lee, and Kim 2013), MEF (Ma, Zeng, and Wang 2015), NPE (Wang et al. 2013), and LIME (Guo, Li, and Ling 2017)) in supplementary materials. Metrics We evaluate the performance of each method using various metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Ablation Study We conduct ablation studies on the LOL-v1 dataset to investigate the effectiveness of our main contributions. |
| Researcher Affiliation | Academia | Yan Huang1, Xiaoshan Liao1, Jinxiu Liang2,3#, Yuhui Quan1,4, Boxin Shi2,3, Yong Xu1,4 1School of Computer Science and Engineering, South China University of Technology, China 2State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, China 3National Engineering Research Center of Visual Technology, School of Computer Science, Peking University, China 4Pazhou Lab, China EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods using mathematical equations and textual explanations, but no structured pseudocode or algorithm blocks are present. |
| Open Source Code | Yes | Code https://github.com/Eileen000/LLIEDiff Extended version https: //github.com/Eileen000/LLIEDiff/raw/main/paper.pdf |
| Open Datasets | Yes | Datasets We conduct experiments on two widely used datasets with paired low-light and normal-light images, LOLv1 (Wei et al. 2018) and LOL-v2 (Yang et al. 2020). We also provide visual comparisons on datasets without ground truth (DICM (Lee, Lee, and Kim 2013), MEF (Ma, Zeng, and Wang 2015), NPE (Wang et al. 2013), and LIME (Guo, Li, and Ling 2017)) in supplementary materials. |
| Dataset Splits | No | The paper mentions using 'LOLv1' and 'LOL-v2' datasets and their respective test sets but does not specify the training/validation/test split percentages, sample counts, or a detailed methodology for how these datasets were partitioned by the authors for their experiments. |
| Hardware Specification | Yes | Our method is implemented using Py Torch and run on a single NVIDIA GTX 3090 Ti GPU. |
| Software Dependencies | No | The paper states, 'Our method is implemented using Py Torch', but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | Implementation Details Our method is implemented using Py Torch and run on a single NVIDIA GTX 3090 Ti GPU. In all experiments, we use DDIM sampling (Song, Meng, and Ermon 2020) with T = 1000 steps. To further enhance local contrast in the degradation model, we apply pixel-wise exposure β(i) = L(i)exp (L(i) φ) adaptive to the illumination map L at each pixel i, where φ is a hyperparameter set to 0.3. The Gaussian blur filter G applied to the bright channel in Eq. (4) uses a kernel size of 5 5 with standard deviations 1.5. To address input with resolutions different from the pretrained model s training resolution, we employ the method proposed in Wang et al. (2023b). All experiments in this section were conducted with inputs of 256 256 resolution to accelerate the sampling process. |