Improving Diffusion Models for Inverse Problems using Manifold Constraints
Authors: Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, Jong Chul Ye
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | With extensive experiments, we show that our method is superior to the previous methods both theoretically and empirically, producing promising results in many applications such as image inpainting, colorization, and sparse-view computed tomography. |
| Researcher Affiliation | Academia | Korea Advanced Institute of Science and Technology (KAIST) |
| Pseudocode | No | The paper describes algorithmic steps and equations (e.g., in Section 3) but does not include a formally labeled "Algorithm" or "Pseudocode" block. |
| Open Source Code | Yes | Code available here |
| Open Datasets | Yes | For inpainting, we use FFHQ 256 256 [24], and Image Net 256 256 [12] to validate our method. ... For the colorization task, we use FFHQ 256 256, and LSUN-bedroom 256 256 [51]. ... For experiments with CT, we train our model based on ncsnpp as a VE-SDE from score-SDE [41], on the 2016 American Association of Physicists in Medicine (AAPM) grand challenge dataset, and we process the data as in [23]. |
| Dataset Splits | Yes | We validate the performance on 1000 held-out validation set images for both FFHQ and Image Net dataset. ... We use 300 validation images for testing the performance with respect to the LSUN-bedroom dataset. ... Evaluation is performed on 421 held-out validation images from the AAPM challenge. |
| Hardware Specification | Yes | Method Wall-clock time [s] ... Computed with a single GTX 1080Ti GPU. |
| Software Dependencies | No | The paper mentions a "Py Torch implementation" and the possibility of switching to a "JAX [6] implementation" but does not provide specific version numbers for any software dependencies used in the experiments. |
| Experiment Setup | Yes | Lastly, we observe the difference in the performance as we vary the values of α. ... Inspecting Fig. 5b, we see that α values within the range [0.1, 1.0] produce satisfactory results. ... As the performance of diffusion models depend heavily on the number of NFEs, we observe the trade-off of each diffusion model when varying the NFE from 20 to 1000. |