Hierarchical Integration Diffusion Model for Realistic Image Deblurring
Authors: Zheng Chen, Yulun Zhang, Ding Liu, bin xia, Jinjin Gu, Linghe Kong, Xin Yuan
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on synthetic and real-world blur datasets demonstrate that our HI-Diff outperforms state-of-the-art methods. |
| Researcher Affiliation | Collaboration | 1Shanghai Jiao Tong University, 2ETH Zürich, 3Bytedance Inc, 4Tsinghua University, 5Shanghai AI Laboratory, 6The University of Sydney, 7Westlake University |
| Pseudocode | No | The paper includes architectural diagrams (Figure 1) but no explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and trained models are available at https://github.com/zhengchen1999/HI-Diff. |
| Open Datasets | Yes | Following previous image deblurring methods, we evaluate our method on synthetic datasets (Go Pro [28] and HIDE [39]) and the real-world dataset (Real Blur [34] and RWBI [56]). |
| Dataset Splits | No | The paper specifies training and testing splits for datasets (e.g., 'Go Pro dataset contains 2,103 pairs of blurry and sharp images for training and 1,111 image pairs for testing.'), but does not explicitly mention a separate validation split or its size. |
| Hardware Specification | Yes | We use Py Torch [31] to implement our models with 4 A100 GPUs. |
| Software Dependencies | No | The paper mentions 'Py Torch [31]' but does not provide a specific version number for PyTorch or any other software dependency. |
| Experiment Setup | Yes | We train our HI-Diff with Adam optimizer [19] with β1=0.9 and β2=0.99. For stage one, the total training iterations are 300K. The initial learning rate is set as 2 10 4 and gradually reduced to 1 10 6 with the cosine annealing [27]. Following previous work [53], we adopt progressive learning. Specifically, we set the initial patch size as 128 and the patch size as 64. We progressively update the patch size and batch size pairs to [(1602,40),(1922,32),(2562,16),(3202,16),(3842,8)] at iterations [20K,40K,60K,80K,100K]. For stage two, we adopt the same training settings as in stage one. Moreover, following previous works [53, 54], we apply random rotation and flips for data augmentation. |