Yanlong Yang
← All projects

Research2025 – 2026

DeblurSDI — Self-Diffusion Driven Blind Imaging

First author · CVPR 2026

Blind deblurring on FFHQ faces. DeblurSDI recovers both the sharp image and the blur kernel (inset) with no pre-training.
Blind deblurring on FFHQ faces. DeblurSDI recovers both the sharp image and the blur kernel (inset) with no pre-training. CVPR 2026 paper
blur kernel, estimated jointlynoise → self-diffusion → sharp image
Schematic

Blind inverse problems must estimate the latent image and the unknown degradation at the same time. Existing methods suffer from unstable convergence, limited priors and sensitivity to hyperparameters.

DeblurSDI treats recovery as an iterative reverse self-diffusion process that starts from pure noise and progressively refines both the sharp image and the blur kernel. It is zero-shot and self-supervised — no pre-training and no external data — and outperforms other methods by a clear margin on combined optical aberrations and motion blur.

Highlights

  • Joint estimation of image and blur kernel
  • No pre-training, no external data
  • Published at CVPR 2026
CVPR paperarXiv

Publication

2026Conference

★Self-Diffusion Driven Blind Imaging

Y. Yang, G. Luo

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · pp. 26711–26720

Zero-shot, self-supervised blind imaging: an iterative reverse self-diffusion process that starts from pure noise and jointly refines the sharp image and the unknown blur kernel — no pre-training.

PaperarXivProject