Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model
Authors: Chunming He, Chengyu Fang, Yulun Zhang, Longxiang Tang, Jinfa Huang, Kai Li, zhenhua guo, Xiu Li, Sina Farsiu
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Reti-Diff outperforms existing methods on three IDIR tasks, as well as downstream applications. |
| Researcher Affiliation | Collaboration | 1Duke University, 2Tsinghua University, 3Shanghai Jiao Tong University, 4Peking University, 5Meta, 6Tianyi Traffic Technology, |
| Pseudocode | No | The paper describes its methodology in Section 3 and illustrates the framework in Figure 3, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at https://github.com/Chunming He/Reti-Diff. |
| Open Datasets | Yes | We conduct experiments on four datasets: LOL-v1 (Wei et al., 2018), LOL-v2-real (Yang et al., 2021), LOL-v2-syn (Yang et al., 2021), and SID (Chen et al., 2019)... We select two widely-used underwater image enhancement datasets: UIEB (Li et al., 2019) and LSUI (Peng et al., 2023)... Following CLIP-LIT (Liang et al., 2023), we select the BAID (Lv et al., 2022) dataset... we selected five commonly-used realworld datasets, i.e., DICM (Lee et al., 2013), LIME (Guo et al., 2016), MEF (Wang et al., 2013), NPE (Ma et al., 2015), and VV (He et al., 2024b)... all compared methods are performed on Ex Dark (Loh & Chan, 2019)... For semantic segmentation, following (Ju et al., 2022), we apply image darkening to samples from the VOC (Everingham et al., 2010) dataset... concealed object segmentation (COS) on two datasets, COD10K (Fan et al., 2021) and NC4K (Lv et al., 2021) |
| Dataset Splits | Yes | Adhering to the training manner in (Cai et al., 2023), we compare our method against 17 cutting-edge techniques and report the results in Table 1. |
| Hardware Specification | Yes | Our Reti-Diff is implemented in Py Torch on four RTX4090 GPUs |
| Software Dependencies | No | The paper mentions "implemented in Py Torch" but does not specify a version number for PyTorch or any other software libraries. |
| Experiment Setup | Yes | Our Reti-Diff is implemented in Py Torch on four RTX4090 GPUs and is optimized by Adam with momentum terms (0.9, 0.999). In phases I and II, we train the network for 300K iterations and the learning rate is initialized as 2 10 4 and gradually reduced to 1 10 6 with the cosine annealing (Loshchilov, 2016). Random rotation and flips are used for augmentation. Reti-Diff comprises RLDM and RGformer. For RLDM, the channel number C and the total time step T are set as 64 and 4. β1:T linearly increase from β1 = 0.1 to βT = 0.99. RGformer adopts a 4-level cascade structure. We set the number of transformer blocks, the attention heads, the channel number as [3, 3, 3, 3], [1, 2, 4, 8], [64, 128, 256, 512] from level 1 to 4. |