TL;DR

SotA Blind-SR methods can super-resolve well LR images which were downscaled by simple, low-pass-filter kernels (e.g., (an)isotropic Gaussians), and fail on more complex downscaling kernels, which are outside their training distribution. In fact, for LR images obtained by non-Gaussian downscaling kernels, SotA Blind-SR methods perform worse than simple interpolation.
We introduce KernelFusion, a zero-shot diffusion-based approach that learns an image-specific patch prior from the input LR image and jointly estimates an unrestricted SR kernel and the HR reconstruction by maximizing cross-scale patch consistency. By breaking free from predefined kernel assumptions and training distributions, KernelFusion establishes a new paradigm of zero-shot Blind-SR that can handle unrestricted, image-specific kernels previously thought impossible.

Our approach consists of 2 stages: Phase 1: We train a diffusion model (PD) to learn the patch distribution of a single image.
Phase 2: We perform blind SR and kernel estimation simultaneously. In particular, we use the trained PD to shift the HR guess toward the patch distribution of the LR input. A refinement U-Net and an implicit kernel representation model are trained jointly under a consistency loss, ensuring that convolving the estimated HR image with the learned kernel reproduces the original LR image.

Our evaluation targeted unrestricted kernels through: (i) artificial complex non-Gaussian kernels, (ii) real measured downscaling kernels[1].
DIV2KFK: DIV2K-val set downscaled using 8 real camera kernels from Levin et al. (2009), induced by small camera jitter during the shutter exposure time. Each image uses a randomly chosen downscaling kernel.
Blind144: 12 images × 12 kernels (144 cases) to isolate the effect of image content vs. kernel.

We further evaluated our method on real-world images from various different sources.

Kernel estimation results on Blind144. The top row displays the 12 ground-truth (GT) degradation kernels, including real-world motion blur kernels from Levin et al. (2009), an anisotropic Gaussian kernel, and three synthetic non-natural kernels (L-shape, empty square, and filled square). The subsequent rows show our method’s estimated kernels for each of the 12 kernels applied to 6 sample images of the DIV2K validation set. Our approach successfully captures a diverse range of degradations, including complex structured kernels, demonstrating its robustness and adaptability in blind SR kernel estimation.

We further evaluate KernelFusion on real-world images from various sources, where the true downscaling kernel is unknown. Notably, the recovered kernels are non-Gaussian, highlighting that real-world degradations deviate from the simplistic assumptions made by existing Blind-SR methods. Our method produces sharp reconstructions without any assumption on the degradation.

References

[1] Netalee Efrat, Daniel Glasner, Alexander Apartsin, Boaz Nadler, and Anat Levin. "Accurate blur models vs. image priors in single image super-resolution." ICCV 2013.

Bibtex

@inproceedings{
	Heinimann2026KernelFusion,
	title={KernelFusion: Zero-Shot Blind Super-Resolution via Patch Diffusion},
	author={Oliver Heinimann, Tal Zimbalist, Assaf Shocher and Michal Irani},
	booktitle={The Fourteenth International Conference on Learning Representations},
	year={2026},
	url={https://arxiv.org/abs/2503.21907}
}