SITCOM: Step-wise Triple-Consistent Diffusion Sampling For Inverse Problems
Authors: Ismail Alkhouri, Shijun Liang, Cheng-Han Huang, Jimmy Dai, Qing Qu, Saiprasad Ravishankar, Rongrong Wang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Compared to SOTA baselines, our experiments across several linear and non-linear tasks (with natural and medical images) demonstrate that SITCOM achieves competitive or superior results in terms of standard similarity metrics and run-time. Extensive Evaluation: We evaluate SITCOM on one medical image reconstruction task (MRI) and 8 image restoration tasks (5 linear and 3 non-linear). Compared to leading baselines, our approach consistently achieves either state-of-the-art or highly competitive quantitative results, while also reducing the number of sampling steps and, consequently, the computational time. Ablation studies are given in Appendix J. |
| Researcher Affiliation | Academia | 1Electrical Engineering & Computer Science at University of Michigan Ann Arbor 2Computational Mathematics, Science, & Engineering at Michigan State University (MSU) 3Biomedical Engineering at MSU 4Mathematical Sciences at MSU. |
| Pseudocode | Yes | Algorithm 1 Step-wise Triple-Consistent Sampling (SITCOM). Input: Measurements y, forward operator A( ), pre-trained DM ϵθ( , ), number of diffusion steps N, DM noise schedule αi for i {1, . . . , N}, number of gradient updates K, stopping criterion δ, learning rate γ, and regularization parameter λ. Output: Restored image ˆx. |
| Open Source Code | Yes | Our code is available online.3 3https://github.com/sjames40/SITCOM |
| Open Datasets | Yes | We use 100 test images from the validation set of FFHQ (Karras et al., 2019) and 100 test images from the validation set of Image Net (Deng et al., 2009) for which the FFHQ-trained and Image Net-trained DMs are given in (Chung et al., 2023b) and (Dhariwal & Nichol, 2021), respectively, following the previous convention. We use the fast MRI dataset (et al, 2019). |
| Dataset Splits | Yes | We use 100 test images from the validation set of FFHQ (Karras et al., 2019) and 100 test images from the validation set of Image Net (Deng et al., 2009)... For testing, we used 50 images taken from the validation dataset. For training the Supervised U-Net (Ronneberger et al., 2015) and E2E Varnet (Sriram et al., 2020), we selected 8,000 training images from the 973 available volumes, omitting the first and last five slices from each volume. |
| Hardware Specification | Yes | All experiments were conducted using a single RTX5000 GPU machine. |
| Software Dependencies | No | The paper mentions software like "ADAM optimizer (Kingma & Ba, 2015)", "BART toolbox (Tamir et al., 2016)", and "U-Net (Ronneberger et al., 2015)", but it does not specify version numbers for these software components or any programming languages or libraries used. |
| Experiment Setup | Yes | For SITCOM, Table 17 in Appendix K.1 lists all the hyper-parameters used for every task. Our experimental setup for image restoration problems and noise levels largely follows DPS (Chung et al., 2023b). The learning rate in Algorithm 1 is set to γ = 0.01 for all tasks, datasets, and measurement noise levels. |