EPA: Boosting Event-based Video Frame Interpolation with Perceptually Aligned Learning

In NIPS 2025.

This repository represents the official PyTorch implementation for the paper "EPA: Boosting Event-based Video Frame Interpolation with Perceptually Aligned Learning", also denoted as "EPA".

We present EPA, a novel framework designed to address a critical challenge in Event-based Video Frame Interpolation (E-VFI): performance degradation in extreme scenarios with high-speed motion and severe keyframe degradation (e.g., blur, noise). The core innovation of EPA is a paradigm shift from conventional pixel-level supervision to learning within a degradation-insensitive, semantic-perceptual feature space.

• Perceptually Aligned Learning Paradigm: By operating in a semantic-perceptual feature space, EPA is significantly more robust to real-world degradations like motion blur and sensor noise, leading to superior generalization.
• Bidirectional Event-Guided Alignment (BEGA) Module: We propose a novel and efficient module that leverages the high temporal resolution of event streams to accurately align and fuse semantic features from keyframes.
• State-of-the-Art Performance: EPA achieves leading performance on multiple synthetic and real-world benchmarks (e.g., GOPRO, Vimeo90k, HS-ERGB), especially in terms of perceptual quality metrics like LPIPS and DISTS.

🛠️ Installation

Setup Environment

We highly recommend using conda to create an isolated environment.

# 1. Clone the repository
git clone https://github.com/your-username/EPA.git
cd EPA

# 2. Create and activate the conda environment
conda create -n epa python=3.10
conda activate epa

# 3. Install dependencies
pip install -r requirements.txt

💾 Datasets and Pre-trained Models

1. Datasets

Please download the required datasets and organize them as follows.

1. Vimeo90k dataset
2. GOPRO dataset
3. HSERGB dataset
4. BSERGB dataset
5. EventAid-F dataset

2. Pre-trained Models

It will be available soon.

⚡️ Inference

To run inference on your own frame sequence, please refer to the interpolation.py file.

📜 Citation

If you find our work useful for your research, please consider citing our paper:

@inproceedings{liu2025epa,
  title={{EPA}: Boosting Event-based Video Frame Interpolation with Perceptually Aligned Learning},
  author={Liu, Yuhan and Fu, Linghui and Yang, Zhen and Chen, Hao and Li, Youfu and Deng, Yongjian},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2025}
}

🙏 Acknowledgements

Our implementation builds upon several excellent open-source projects. We are grateful for their contributions to the community.

• DINO
• PerVFI
• FloLPIPS

📄 License

This project is licensed under the MIT License.