Restabilizing Diffusion Models with Predictive Noise Fusion Strategy for Image Super-Resolution
Authors: Luoqian Jiang, Yong Guo, Bingna Xu, Haolin Pan, Jiezhang Cao, Wenbo Li, Jian Chen
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that PNFS significantly improves the stability and performance of diffusion models in super-resolution, both quantitatively and qualitatively. Furthermore, PNFS can be flexibly integrated into various diffusion models to enhance their stability. ... We conducted experiments with a set of 10 Gaussian noises. ... Experiments on multiple benchmark datasets demonstrate that, our PNFS yields significantly superior results both quantitatively and qualitatively, consistently producing high-quality images. |
| Researcher Affiliation | Academia | 1South China University of Technology 2Harvard University 3The Chinese University of Hong Kong EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Scheme of the Predictive Noise Fusion Strategy (PNFS). |
| Open Source Code | Yes | Code https://github.com/Rosiekk/PNFS-main |
| Open Datasets | Yes | The training and validation sets consist of 900 and 100 images randomly selected from the DIV2K and Flickr2K (Agustsson and Timofte 2017) datasets. ... We evaluate different methods on well-known SR datasets, including Set5 (Bevilacqua et al. 2012), Set14 (Zeyde, Elad, and Protter 2010), BSD100 (Martin et al. 2001), Urban100 (Huang, Singh, and Ahuja 2015), Manga109 (Matsui et al. 2017) and Real SR (Cai et al. 2019). |
| Dataset Splits | Yes | The training and validation sets consist of 900 and 100 images randomly selected from the DIV2K and Flickr2K (Agustsson and Timofte 2017) datasets. For each image, 100 Gaussian noises are sampled to generate 90,000 training samples. Random crops of size 512 512 are used for training, and evaluation is performed on full-size images. |
| Hardware Specification | Yes | All the experiments are conducted on a server with NVIDIA Ge Force RTX 3090 24G GPU. |
| Software Dependencies | No | The paper mentions 'Adam optimizer' and 'DDIM' as a sampling strategy but does not provide specific version numbers for these or other software libraries (e.g., Python, PyTorch/TensorFlow) that would be needed for replication. |
| Experiment Setup | Yes | DPM is trainable and trained for 160K iterations with a batch size of 100. We use the Adam optimizer with an initial learning rate of 6e 5, adjusted using a poly learning rate schedule with a default factor of 1.0. The training and validation sets consist of 900 and 100 images randomly selected from the DIV2K and Flickr2K (Agustsson and Timofte 2017) datasets. For each image, 100 Gaussian noises are sampled to generate 90,000 training samples. Random crops of size 512 512 are used for training, and evaluation is performed on full-size images. During SR task inference, we freeze all parameters of the pre-trained diffusion model and the DPM, using DDIM (Song, Meng, and Ermon 2020) as the sampling strategy for all experiments. |