Towards Realistic Data Generation for Real-World Super-Resolution
Authors: Long Peng, Wenbo Li, Renjing Pei, Jingjing Ren, Jiaqi Xu, Yang Wang, Yang Cao, Zheng-Jun Zha
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Real DGen excels in generating large-scale, high-quality paired data that mirrors real-world degradations, significantly advancing the performance of popular SR models on various real-world benchmarks. Extensive experiments show that Real DGen outperforms previous methods in generating realistic paired data and enhancing the performance of SR models in real-world scenarios. We conduct an ablation experiment. |
| Researcher Affiliation | Collaboration | 1 USTC, 2 Huawei Noah s Ark Lab, 3 HKUST(GZ), 4 CUHK, 5 Chang an University {longp2001@mail.,ywang120@}ustc.edu.cn,EMAIL |
| Pseudocode | Yes | Algorithm 1 Decoupled DDPM Training; Algorithm 2 Data Generation |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | We collect about 152,000 real low-resolution images from both public datasets (Wei et al., 2020; Cai et al., 2019; Ignatov et al., 2017) and those captured using smartphones to train Real DGen. To ensure a fair comparison, we compare our approach with existing methods by using the widely-used DIV2K dataset (Agustsson & Timofte, 2017) as HR images and real LR images as degradation references to create the paired training data for training various popular SR models. |
| Dataset Splits | Yes | We collect about 152,000 real low-resolution images from both public datasets (Wei et al., 2020; Cai et al., 2019; Ignatov et al., 2017) and those captured using smartphones to train Real DGen. To further explore the performance comparison on real paired data, we further train the RRDB model using the real-collected training subset of Real SR and evaluate it on the test subset, as shown in Table 14. |
| Hardware Specification | Yes | Extractors and Decouple DDPM are trained with learning rate 1 10 4 and batch size 16 on 16 NVIDIA V100 GPUs. We utilize the public Basic SR for training Real-SR methods with 16 NVIDIA V100 GPUs. Our fine-tuning regimen is conducted with a learning rate of 1 10 5, utilizing 8 NVIDIA V100 GPUs, which spent approximately one week for training. The training is executed on 8 NVIDIA V100 GPUs over the course of approximately seven days, with the learning rate configured at 1e 4. The batch size is defined as 8, and the entire training regimen is on 8 NVIDIA V100 GPUs, which typically consume around 14 days to complete. Using the official implementation, we measure inference time on 4 V100 GPUs. |
| Software Dependencies | No | The paper mentions 'public Basic SR for training Real-SR methods' but does not specify a version number for Basic SR or any other software dependency. |
| Experiment Setup | Yes | Extractors and Decouple DDPM are trained with learning rate 1 10 4 and batch size 16 on 16 NVIDIA V100 GPUs. We utilize L1Loss (Chen et al., 2023b; Liang et al., 2021) and perception GAN loss (Wang et al., 2021b; Johnson et al., 2016; Wang et al., 2018) for training PNSR-oriented and Perception-oriented Real SR models, respectively. During the training of the Decoupled DDPM, we configure the maximum diffusion step to be 500, span the training over 100 epochs, and utilize a learning rate of 1e 4 for the decoupled diffusion model. For the fine-tuning of the extractors, we apply a more refined learning rate of 1e 6. The batch size is defined as 8. In the loss function of the manuscript, we set the number of samples n and margin to 3 and 0.01, respectively, based on our experimental results. |