DiffFPR: Diffusion Prior for Oversampled Fourier Phase Retrieval
Authors: Ji Li, Chao Wang
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct experiments to evaluate the performance of our approach on Fourier phase retrieval. Two datasets, including Flickr Faces High Quality (FFHQ) 256 256 and Image Net 256 256 are considered. The pretrained diffusion models are directly downloaded from the open-source library without any refinement. Each test dataset contains 1K images. To quantitatively compare the performance, we report the reference-based PSNR/SSIM metric to measure the closeness to the original image, and the LPIPS metric to measure the perception quality of the restoration. |
| Researcher Affiliation | Academia | 1Academy for Multidisciplinary Studies, Capital Normal University, Beijing, China 2University of Kansas Medical Center, Kansas City, US. |
| Pseudocode | Yes | Algorithm 1 RAAR for Fourier phase retrieval. Algorithm 2 Measurement-guided diffusion model for Fourier phase retrieval. |
| Open Source Code | Yes | The code is available at https: //github.com/Chilie/Diff FPR. |
| Open Datasets | Yes | Two datasets, including Flickr Faces High Quality (FFHQ) 256 256 and Image Net 256 256 are considered. |
| Dataset Splits | No | The paper mentions 'Each test dataset contains 1K images' and 'the paired training dataset contains 1000 images' but does not specify a separate validation split or explicit percentages for training, validation, and test sets. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments, such as CPU/GPU models or memory. |
| Software Dependencies | No | The paper mentions 'pretrained diffusion models are directly downloaded from the open-source library' but does not specify any software names with version numbers or other software dependencies. |
| Experiment Setup | Yes | In the default configuration, we select 1000 diffusion-step DDIM as the unconditional sampling, and the one-step RAAR as the inner iterative engine. The hyperparameter β is set to unity and 0.75 for noiseless and noisy cases respectively. ... For all experiments, the oversampling ratio is set to 4. |