Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights
Authors: Shijun Liang, Ismail Alkhouri, Siddhant Gautam, Qing Qu, Saiprasad Ravishankar
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate UGo DIT on both medical (multi-coil MRI) and natural (super resolution and nonlinear deblurring) image recovery tasks under various settings. Compared to recent standalone DIP methods, UGo DIT provides accelerated convergence and notable improvement in reconstruction quality. Furthermore, UGo DIT achieves performance competitive with SOTA DM-based and supervised approaches, despite not requiring large amounts of clean training data. Our code is available at: UGo DIT. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering & Computer Science, University of Michigan Ann Arbor 2Department of Computational Mathematics, Science, & Engineering, Michigan State University 3Department of Biomedical Engineering, Michigan State University EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Training: Unsupervised Group DIP Transferable via weights (UGo DIT) Algorithm 2 Testing: Unsupervised Group DIP Transferable via weights (UGo DIT-M) |
| Open Source Code | Yes | Our code is available at: UGo DIT. |
| Open Datasets | Yes | For MRI, we use the knee portion of the fast MRI dataset [22]. For the tasks of SR and NDB, we use the FFHQ dataset [51]. ... For SR, we observe that UGo DIT achieves slightly better PSNR than a Seq DIP and eventually converges to similar PSNR values. However, unlike the MRI case, UGo DIT-OOD converges faster than a Seq DIP. This can also be observed in the restored images of Figure 8 and Figure 9. We attribute UGo DIT s ability to generalize well to learning multi-frequency features under OOD scenario for the natural image. In Table 3, we report test-time average PSNR on 20 CBSD688 images |
| Dataset Splits | Yes | For all three tasks, we test UGo DIT and baselines with 20 randomly selected degraded images6. ... From each volume, we discard the first and last five slices, then randomly sample 8,000 images from the remaining slices to form the training set. |
| Hardware Specification | Yes | All the experiments are run on a single RTX5000 GPU machine. |
| Software Dependencies | No | The paper mentions "ADAM optimizer [47]" but does not specify a version number for it or any other software libraries or frameworks used in the implementation. |
| Experiment Setup | Yes | Step size is set to β = 0.0001. The regularization parameter is set to λ = 2, following the study in Appendix G (where we also show the robustness to noise overfitting). For (N, K), we use (2, 2000), (10, 2000), and (10, 2000) for MRI, SR, and NDB, respectively, following the study in Appendix M. |