Yanlong Yang
← All projects

Research2025

Self-Diffusion for Solving Inverse Problems

Co-author · arXiv 2025

One untrained model, many inverse problems: inpainting, motion deblurring, 4× super-resolution, and accelerated 2D and 3D MRI.
One untrained model, many inverse problems: inpainting, motion deblurring, 4× super-resolution, and accelerated 2D and 3D MRI. arXiv 2025 paper
noise ⇄ denoise, untrained network×T steps
Schematic

Self-diffusion alternates noising and denoising with an untrained convolutional self-denoiser. Because nothing is pretrained, the same procedure adapts to arbitrary forward operators — a general recipe for inverse problems where clean training data is scarce.

Highlights

  • Untrained convolutional self-denoiser
  • Works with arbitrary forward operators
  • Foundation for DeblurSDI
arXiv

Publication

2025Preprint

Self-Diffusion for Solving Inverse Problems

G. Luo, S. Huang, Y. Yang

arXiv:2510.21417

Alternates noising and denoising with an untrained convolutional self-denoiser, adapting to arbitrary forward operators without pretrained priors.

arXivProject