[CVPR 2025 Oral] Diffusion Fourier Neural Operator for Arbitrary-Scale Super-Resolution

Xiaoyi Liu1, Hao Tang2
1Washington University in St. Louis  |  2Peking University

PSNR vs Inference Time

Abstract

We introduce DiffFNO, a novel diffusion framework for arbitrary-scale super-resolution strengthened by a Weighted Fourier Neural Operator (WFNO). Mode Rebalancing in WFNO effectively captures critical frequency components, significantly improving the reconstruction of high-frequency image details. Gated Fusion Mechanism (GFM) adaptively complements WFNO’s spectral features with spatial features from an Attention-based Neural Operator (AttnNO). Adaptive Time-Step (ATS) ODE solver accelerates inference without sacrificing output quality. Extensive experiments demonstrate that DiffFNO achieves state-of-the-art results, outperforming existing methods by 2–4 dB in PSNR across various scaling factors and generalizing beyond training distributions—all at competitive inference times.

Architecture

DiffFNO architecture

Architecture diagram.

Qualitative Results

Qualitative results

Sharper edges, richer textures.

Quantitative Results

PSNR and SSIM results table

PSNR/SSIM on the DIV2K validation set across scales ×2–×12.

BibTeX

@inproceedings{liu2025difffno,
  title={DiffFNO: Diffusion Fourier Neural Operator},
  author={Liu, Xiaoyi and Tang, Hao},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={150--160},
  year={2025}
}