UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration
Authors: Minghao Liu, Wenhan Yang, Jinyi Luo, Jiaying Liu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our method achieves significant performance improvements across various image restoration tasks, realizing true all-in-one image restoration. To demonstrate the effectiveness of our proposed method, we evaluate it on three primary image restoration tasks: image dehazing, image deraining, and image denoising. [...] The results are presented in Table 1. [...] In the ablation study, we assess the impact of incorporating Prompt Blocks at different levels of our model... |
| Researcher Affiliation | Academia | 1Wangxuan Institute of Computer Technology, Peking University 2Peng Cheng Laboratory EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the optimization problem and its solution using ADMM in textual form and mathematical equations, but does not present a distinct pseudocode block or algorithm listing. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code, nor does it provide a link to a code repository. |
| Open Datasets | Yes | For these tasks, we use datasets: BSD400, BSD68 (Martin et al. 2001), and WED (Ma et al. 2016) for image denoising, Rain100L (Yang et al. 2020) for image deraining, and RESIDE (Li et al. 2018a) for image dehazing. [...] For deblurring, we used the Go Pro dataset; for desnowing, we selected 611 images from the Snow100K dataset; and for super-resolution, we used the Real SR dataset. [...] For HFD-IQA, we trained the required SVR model using degraded images from the TID2013 dataset, with their NQM metrics as features extracted by a pre-trained Res Net50 model. For Hyper-IQA, we utilize the pre-trained model on the Koniq-10k dataset, as provided by the authors. |
| Dataset Splits | Yes | Following the division in (Li et al. 2022), BSD400 and WED are used for training while BSD68 is used for testing with 68 ground truth images. For image deraining, we use the 1800 rain-clean paired images and 100 testing pairs provided in the Rain100L dataset. For image dehazing, we use the Outdoor Training Set (OTS) for training and the Synthetic Objective Testing Set (SOTS) for testing from the RESIDE dataset. |
| Hardware Specification | Yes | Training the model to convergence requires only one day on a single 4090 GPU, using 128 × 128 cropped patches as input. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and references C2F-DFT for the backbone model but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | Our framework consists of a four-level encoder-decoder structure, with each level’s Diffusion Transformer containing different numbers of Transformer blocks: [4, 6, 6, 8] from the first to the fourth level. We use a Prompt Block between every two consecutive decoder levels, totaling three Prompt Blocks in the network. The model is trained in a multi-degradation all-in-one setting with batch sizes of 120 and 20 for the two training phases, and a batch size of 120 in the single degradation setting. The network is optimized using the Adam optimizer (β1=0.9, β2=0.999), with the learning rate reduced to 0.01 after 50,000 epochs. Training the model to convergence requires only one day on a single 4090 GPU, using 128 × 128 cropped patches as input. |