Improving Diffusion Models for Inverse Problems Using Optimal Posterior Covariance
Authors: Xinyu Peng, Ziyang Zheng, Wenrui Dai, Nuoqian Xiao, Chenglin Li, Junni Zou, Hongkai Xiong
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that the proposed methods significantly enhance reconstruction performance without requiring hyperparameter tuning. |
| Researcher Affiliation | Academia | 1School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China. Correspondence to: Ziyang Zheng <zhengziyang@sjtu.edu.cn>, Wenrui Dai <daiwenrui@sjtu.edu.cn>, Junni Zou <zoujunni@sjtu.edu.cn>. |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | The source code is available at https://github.com/xypeng9903/k-diffusion-inverse-problems. |
| Open Datasets | Yes | Following (Chung et al., 2023a; Wang et al., 2023), we perform experiments on the FFHQ 256x256 and Image Net 256x256 datasets to compare different methods with unconditional diffusion models from (Chung et al., 2023a) and (Dhariwal & Nichol, 2021), respectively. |
| Dataset Splits | No | The paper mentions using 'FFHQ 256x256' and 'Image Net 256x256' datasets but does not explicitly provide specific train/validation/test split percentages, sample counts, or detailed splitting methodology. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, processor types, memory amounts) used for running experiments were mentioned in the paper. |
| Software Dependencies | No | The paper mentions 'k-diffusion' and 'scipy.sparse.linalg.cg' but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | All measurements are corrupted by Gaussian noise with σ = 0.05. To evaluate different methods, we follow Chung et al. (2023a) to use three metrics: Structure Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS, Zhang et al. (2018)) and Frechet Inception Distance (FID, Heusel et al. (2017)). For the sampler setup, all Type I methods use the same Heun's 2nd deterministic sampler suggested in Karras et al. (2022) with 50 sampling steps, and all Type II methods use the same Heun's 2nd stochastic sampler (Schurn = 80, Stmin = 0.05, Stmax = 50, Snoise = 1.003, definition see Karras et al. (2022)) with 50 sampling steps. |