HIFER-Mamba: High-frequency Information Feature Enhancement and Reconstruction for Image Super-Resolution
Official implementation of HIFER-Mamba, a frequency-aware Mamba framework for image super-resolution.
- 🔥 2026.07 Paper submitted to IEEE Transactions on Image Processing (TIP) and currently under review.
- 🔥 Project page, pretrained models and training code will be released after paper acceptance.
HIFER-Mamba is built upon MambaIR and introduces three lightweight yet effective improvements for image super-resolution:
- Frequency-aware Feature Enhancement
- Progressive Feature Integration
- Nonlinear State Enhancement
In addition, we highlight There specific innovations below:
-
** Frequency-Enhanced State Space Layer (F)**: To enhance the global context modeling capability of Mamba, we introduce a frequency-enhanced state space layer, as illustrated in Fig. 2. Unlike conventional spatial-domain feature propagation, the proposed design explicitly models spectral information to better capture global structures and high-frequency details.
-
Dynamic Prompt with Cross-Layer Consistency (P): To provide adaptive guidance for different layers, we introduce a dynamic prompt generation mechanism. Unlike conventional static prompts that remain fixed during inference, our prompts are dynamically generated according to the current input features, enabling more flexible and content-aware feature modulation across different image structures and degradation patterns.
-
Nonlinear State Expansion (U): Compared with conventional linear state transitions, the proposed nonlinear expansion mechanism enables more expressive feature interaction and improves the modeling of complex spatial structures. In particular, the gated nonlinear transformation enhances feature selectivity and stabilizes deep feature propagation, which is beneficial for recovering fine textures and structurally rich regions in image super-resolution tasks.
These complementary designs significantly improve high-frequency texture reconstruction, structural recovery and feature representation while preserving the efficiency of state-space models.
Image super-resolution requires recovering fine textures while maintaining structural consistency.
HIFER-Mamba integrates frequency-domain representation, progressive feature interaction and nonlinear state modeling into a unified Mamba framework, leading to more faithful image reconstruction.
HIFER-Mamba consistently achieves competitive quantitative performance on five widely used benchmark datasets.
Our model reconstructs sharper edges, richer textures and more faithful structural details than previous Mamba-based methods.
The training process exhibits smooth convergence and stable optimization throughout learning.
- ✅ Frequency-aware enhancement for recovering high-frequency textures.
- ✅ Progressive feature interaction across deep Mamba blocks.
- ✅ Nonlinear state enhancement for stronger representation capability.
- ✅ Superior quantitative performance on standard SR benchmarks.
- ✅ Stable optimization with efficient inference.
analysis/ Analysis scripts
basicsr/ BasicSR framework
datasets/ Training and testing datasets
experiments/ Configurations and checkpoints
figures/ Figures used in the paper
results/ Reconstruction results
options/ Training/testing options
docs/ Project webpage
python basicsr/train.py -opt options/train/train_HIFER_Mamba_x2.ymlpython basicsr/test.py -opt options/test/test_HIFER_Mamba_x2.yml- Frequency-aware representation improves texture reconstruction.
- Progressive feature integration enhances feature consistency.
- Nonlinear state enhancement increases representation capability.
- Better visual quality with competitive PSNR/SSIM.
- Stable convergence during optimization.
🚧 Coming Soon.
The project webpage, pretrained models, checkpoints and supplementary materials will be released after paper acceptance.
If you find this work useful, please consider citing
@article{jiang2026hifermamba,
title={HIFER-Mamba: High-frequency Information Feature Enhancement and Reconstruction for Image Super-Resolution},
author={Nan Jiang and others},
journal={Under Review},
year={2026}
}This project is built upon the excellent open-source projects BasicSR and MambaIR. We sincerely thank the original authors for making their valuable code publicly available.