Realistic Noise Synthesis with Diffusion Models
Authors: Qi Wu, Mingyan Han, Ting Jiang, Chengzhi Jiang, Jinting Luo, Man Jiang, Haoqiang Fan, Shuaicheng Liu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our RNSD method significantly outperforms existing techniques in synthesizing realistic noise under multiple metrics and improving image denoising performance. |
| Researcher Affiliation | Collaboration | Qi Wu1*, Mingyan Han1*, Ting Jiang1, Chengzhi Jiang1, Jinting Luo1, Man Jiang1, Haoqiang Fan1, Shuaicheng Liu2 1Megvii Technology Inc. 2University of Electronic Science and Technology of China EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: RNSD Training Algorithm 2: RNSD Sampling(DIPS) |
| Open Source Code | Yes | Code https://github.com/wuqi-coder/RNSD |
| Open Datasets | Yes | SIDD: The SIDD dataset (Abdelhamed, Lin, and Brown 2018), utilized for training and evaluation, includes subsets such as SIDD small with 160 image pairs from 5 smartphone cameras, and SIDD medium with double the noise sampling. DND: DND benchmark (Plotz and Roth 2017) provides 50 reference images and their realistic noisy counterparts generated using accurate sensor noise models. LSDIR: LSDIR dataset (Li et al. 2023) contains 84,991 high-quality clean samples. |
| Dataset Splits | Yes | SIDD: The SIDD dataset (Abdelhamed, Lin, and Brown 2018), utilized for training and evaluation, includes subsets such as SIDD small with 160 image pairs from 5 smartphone cameras, and SIDD medium with double the noise sampling. The SIDD validation set (1280 patches from unseen sensor settings) is used to test noise model and denoising models. We assessed RNSD s performance with limited data firstly, training models on 10, 20, and 80 pairs from SIDD Small. |
| Hardware Specification | Yes | Models are trained on an NVIDIA Ge Force RTX 2080 Ti GPU for 2 105 iterations. For testing, we use Exponential Moving Average Decay (EMA) with decay 0.995. Our RNSD model achieves an inference time of 0.15 seconds to generate a batch of 16 128 128 image patches using 5 sampling steps with DIPS on an NVIDIA Ge Force RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions DDPM, Adam optimizer, and Dn CNN network, but does not provide specific version numbers for software libraries or frameworks like PyTorch, TensorFlow, CUDA, or Python. |
| Experiment Setup | Yes | We train the DDPM (Ho, Jain, and Abbeel 2020) diffusion model of our noise generation system with 1000 steps, a gradient accumulation step size of 2, and Adam optimizer (lr = 8 10 5). Training samples are 128 128 crops from original images, with a batch size of 16. Models are trained on an NVIDIA Ge Force RTX 2080 Ti GPU for 2 105 iterations. For testing, we use Exponential Moving Average Decay (EMA) with decay 0.995. When training the denoising networks from scratch, we largely keep the original training hyper-parameters consistent with the baseline denoising models. For finetuning, we reduced the learning rate to 1 10 6, maintaining other parameters, and finetuned for 1 106 iterations before evaluation. |