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..
Self-diffusion for Solving Inverse Problems
Authors: Guanxiong Luo, Shoujin Huang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of our approach on a variety of linear inverse problems, showing that self-diffusion achieves competitive or superior performance compared to other methods. 3 Simulation We apply the self-diffusion inference to recover a sparse 1D signal from compressed sensing measurements. 4 Application To demonstrate the versatility and generality of self-diffusion across a range of inverse problems spanning both natural and medical imaging domains, we evaluate our method on several representative tasks, including 2D and 3D MRI reconstruction, low-level vision restoration (inpainting, deblurring, denoising, super-resolution), and radar angle estimation recovery. |
| Researcher Affiliation | Academia | Guanxiong Luo EMAIL Shoujin Huang EMAIL The paper only provides the authors' names and personal email addresses (gmail.com) without specifying any institutional names, departments, cities, or countries. Therefore, a definitive classification as academia, industry, or collaboration based on the provided text is not possible according to the specified criteria. As a best-effort classification for a research paper at NeurIPS, it is categorized as academia, but this is an inference due to lack of explicit information. |
| Pseudocode | Yes | Algorithm 1 Self-diffusion (SDI) for solving the inverse problem 1: Input: x True T ; noise steps T; iterations-K; learning rate η; forward operator A; initialize θT,0; default noise schedule βstart = 0.0001, βend = 0.01 2: for t = T 1 to 0 do 3: Sample ϵt N(0, I), xt = (xtrue t + σ(t)ϵt) 4: for k = 0 to K 1 do 5: Compute loss Lt,k and θt,k Lt,k 6: Update the weights using optimizer and learning rate η 7: end for 8: Set θt = θt,K 9: Compute xtrue t 1 = Dθt(xt) 10: end for 11: return xtrue 0 |
| Open Source Code | Yes | The code is available at github:ggluo/self-diffusion. |
| Open Datasets | Yes | From the fast MRI validation set [22], we randomly selected 20 samples for each contrast (T1, T1 post contrast, T2, and FLAIR) to form a test set of 80 subjects. We leverage a large-scale autonomous driving perception RADIal dataset [27]. The experiments are performed on 1000 images that are from the dataset Image Net and resized to 256 256. We included denoising results on CBSD68 dataset [28] with noise level σ = 25. |
| Dataset Splits | Yes | From the fast MRI validation set [22], we randomly selected 20 samples for each contrast (T1, T1 post contrast, T2, and FLAIR) to form a test set of 80 subjects. We consider a 12-coil 3D T1-weighted brain scan and apply a Cartesian undersampling mask on the k-space that has a dimension of 160 160 128. The undersampling mask is generated using BART s command poisson with an acceleration factor of 2.5 along two phase-encoding directions. This leads to a total acceleration factor of 8. |
| Hardware Specification | Yes | Table 6: Runtime and iteration comparison for different reconstruction methods. Task, Size ... GPU info MRI, 320 320 ... RTX 4090, 24GB MRI, 320 320 ... RTX 3090 Ti, 24GB MRI, 320 320 ... RTX 4090, 24GB MRI, 320 320 ... RTX 3090 Ti, 24GB low-level vision, 256 256 ... RTX 4090, 24GB low-level vision, 256 256 ... RTX 4090, 24GB inpainting, 5000 3000 ... RTX A6000, 48GB |
| Software Dependencies | No | The paper mentions 'BART's ecalib command' and 'Adam optimizer' and 'code from Git:Levi Borodenko/motionblur' but does not specify any software versions for these or other dependencies. For example, it does not state a Python version or specific library versions like PyTorch, TensorFlow, or CUDA versions. |
| Experiment Setup | Yes | The parameters for Algorithm 1 is {T = 40; K = 200; η = 1e 5; βstart = 4e 3; βend = 1e 6} and Adam optimizer is used. For self-diffusion, we initialize the 2D U-Net with Gaussian initialization {µ = 0, σ = 0.02} for convolutional kernels and zero initialization for biases. We use Adam optimizer in Algorithm 1 and set parameters {T = 40; K = 50; η = 0.001; βstart = 1e 3; βend = 1e 4}. The parameters for Algorithm 1 is {T = 30; K = 100; η = 1e 3; βstart = 2e 2; βend = 1e 4} and Adam optimizer is used. We use the Algorithm 1 for all the tasks above and set it with parameters {T = 40; K = 150; η = 0.001, βstart = 1e 4; βend = 1e 2}. |