FMA-Net++ (ECCV 2026)

   Geunhyuk Youk¹  Jihyong Oh^{† 2}  Munchurl Kim^{† 1}

¹Korea Advanced Institute of Science and Technology (KAIST), South Korea

²Chung-Ang University, South Korea

^†Co-corresponding authors

---

This repository is the official implementation of "FMA-Net++: Motion- and Exposure-Aware Joint Video Super-Resolution and Deblurring".

demo.mp4

👆 Experience User-Interactive Comparisons: Please visit our Project Page to explore more results.

📧 News

• June 18, 2026: FMA-Net++ is accepted to ECCV 2026 🎉
• Dec 04, 2025: This repository is created.

🧬 Previous Work

FMA-Net++ builds upon our previous work, FMA-Net (CVPR 2024), addressing its limitations in handling dynamic exposure and limited temporal receptive fields.

📖 Abstract

Joint video super-resolution and deblurring (VSRDB) aims to restore sharp, HR videos from blurry, LR inputs. A key difficulty is that the exposure duration often varies across frames, changing the extent of motion blur throughout a video. Most existing methods assume a fixed exposure and rely on sliding-window or recurrent designs, which struggle to efficiently capture long-range temporal context under such frame-wise exposure variation.

We present FMA-Net++, a non-recurrent, sequence-level framework built from Hierarchical Refinement with Bidirectional Aggregation (HRBA) blocks that process frames in parallel while hierarchically expanding the temporal receptive field. To handle exposure-dependent blur, an Exposure Time-aware Modulation (ETM) layer conditions features on per-frame exposure embeddings from an Exposure Time-aware Feature Extractor (ETE), guiding an exposure-aware dynamic filtering module to estimate motion- and exposure-aware degradation kernels. Trained solely on synthetic data, FMA-Net++ achieves state-of-the-art accuracy and temporal consistency on our proposed REDS-ME and REDS-RE benchmarks, and generalizes well to GoPro and challenging real-world videos.

🖼️ Method Overview

FMA-Net++ utilizes HRBA blocks for efficient temporal modeling and ETM layers to explicitly handle dynamic exposure changes.

🚀 Code Release Plan

The full code and pretrained models will be released soon.

• Inference code
• Pretrained models
• Training scripts
• Dataset generation scripts

📑 Citation

If you find FMA-Net++ useful, please consider citing:

@inproceedings{youk2026fmanetpp,
    author    = {Youk, Geunhyuk and Oh, Jihyong and Kim, Munchurl},
    title     = {FMA-Net++: Motion- and Exposure-Aware Joint Video Super-Resolution and Deblurring},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year      = {2026}
}

📬 Contact

For any questions, please contact rmsgurkjg@kaist.ac.kr via email.