Image Restoration Through Generalized Ornstein-Uhlenbeck Bridge
Authors: Conghan Yue, Zhengwei Peng, Junlong Ma, Shiyan Du, Pengxu Wei, Dongyu Zhang
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental outcomes showcase the state-of-the-art performance achieved by both models across diverse tasks, including inpainting, deraining, and super-resolution. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Sun Yat-sen University, Guangzhou, Guangdong, China. 2Pengcheng Laboratory. Correspondence to: Dongyu Zhang <zhangdy27@mail.sysu.edu.cn>. |
| Pseudocode | No | The paper describes the method using mathematical equations and prose but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https: //github.com/Hammour-steak/GOUB. |
| Open Datasets | Yes | We have selected the Celeb A-HQ 256 256 datasets (Karras et al., 2018) for both training and testing with 100 thin masks. We have selected the Rain100H datasets (Yang et al., 2017) for our training and testing, which includes 1800 pairs of training data and 100 images for testing. We conducted training and evaluation on the DIV2K validation set for 4 upscaling (Agustsson & Timofte, 2017) |
| Dataset Splits | Yes | We have selected the Celeb A-HQ 256 256 datasets (Karras et al., 2018) for both training and testing with 100 thin masks. We have selected the Rain100H datasets (Yang et al., 2017) for our training and testing, which includes 1800 pairs of training data and 100 images for testing. We conducted training and evaluation on the DIV2K validation set for 4 upscaling (Agustsson & Timofte, 2017) |
| Hardware Specification | Yes | Our models are trained on a single 3090 GPU with 24GB memory for about 2.5 days. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' but does not specify software dependencies like programming languages, libraries, or frameworks with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For all experiments, we use the same noise network, with the network architecture and mainly training parameters consistent with the paper (Luo et al., 2023a). ... The steady variance level λ2 was set to 30 (over 255), and the sampling step number T was set to 100. In the training process, we set the patch size = 128 with batch size = 8 and use Adam (Kingma & Ba, 2015) optimizer with parameters β1 = 0.9 and β2 = 0.99. The total training steps are 900 thousand with the initial learning rate set to 10 4, and it decays by half at iterations 300, 500, 600, and 700 thousand. For the setting of θt, we employ a flipped version of cosine noise schedule (Nichol & Dhariwal, 2021), enabling θt to change from 0 to 1 over time. Notably, to address the issue of θt being too smooth when t closed to 1, we let the coefficient e θT to be a small enough value δ = 0.005 instead of zero, which represents θT PT i=0 θidt = log δ, as well as dt = log δ/ PT i=0 θi. |