VFIMamba: Video Frame Interpolation with State Space Models arxiv

VFIMamba: Video Frame Interpolation with State Space Models
Guozhen Zhang, Chunxu Liu, Yutao Cui, Xiaotong Zhao, Kai Ma, Limin Wang

💥 News

• [2024.09.3] Support for directly importing model weights from HuggingFace. Thanks to the HuggingFace team for their efforts!
• [2024.07.3] Demo and evaluation codes released.

😆 HighLights

In this work, we have introduced VFIMamba, the first approach to adapt the SSM model to the video frame interpolation task. We devise the Mixed-SSM Block (MSB) for efficient inter-frame modeling using S6. We also explore various rearrangement methods to convert two frames into a sequence, discovering that interleaved rearrangement is more suitable for VFI tasks. Additionally, we propose a curriculum learning strategy to further leverage the potential of the S6 model. Experimental results demonstrate that VFIMamba achieves the state-of-the-art performance across various datasets, in particular highlighting the potential of the SSM model for VFI tasks with high resolution.

💕Installation

CUDA 11.7

• torch 1.13.1
• python 3.10.6
• causal_conv1d 1.0.0
• mamba_ssm 1.0.1
• skimage 0.19.2
• numpy
• opencv-python
• timm
• tqdm
• tensorboard

😎 Play with Demos

1. Download the model checkpoints and put the ckpt folder into the root dir. We also support directly importing model weights from HuggingFace. Please refer to hf_demo_2x.py.
2. Run the following commands to generate 2x and Nx (arbitrary) frame interpolation demos:

We provide two models, an efficient version (VFIMamba-S) and a stronger one (VFIMamba). You can choose what you need by changing the parameter model.

Hugging Face Demo

For Hugging Face demo, please refer to the code here.

python hf_demo_2x.py --model **model[VFIMamba_S/VFIMamba]**      # for 2x interpolation

Manually Load

python demo_2x.py  --model **model[VFIMamba_S/VFIMamba]**      # for 2x interpolation
python demo_Nx.py --n 8 --model **model[VFIMamba_S/VFIMamba]** # for 8x interpolation

By running above commands with model VFIMamba, you should get the follow examples by default:

You can also use the scale parameter to improve performance at higher resolutions; We will downsample to scale*shape to predict the optical flow and then resize to the original size to perform the other operations. We recommend setting the scale to 0.5 for 2K frames and 0.25 for 4K frames.

python demo_2x.py  --model VFIMamba --scale 0.5 # for 2K inputs with VFIMamba   

🏃 Evaluation

1. Download the dataset you need:

   • Vimeo90K dataset
   • UCF101 dataset
   • Xiph dataset
   • SNU-FILM dataset
   • X4K1000FPS dataset

2. Download the model checkpoints and put the ckpt folder into the root dir. We also support directly importing model weights from HuggingFace. Please refer to hf_demo_2x.py.

For all benchmarks:

python benchmark/**dataset**.py --model **model[VFIMamba_S/VFIMamba]** --path /where/is/your/**dataset**

You can also test the inference time of our methods on the $H\times W$ image with the following command:

python benchmark/TimeTest.py --model **model[VFIMamba_S/VFIMamba]** --H **SIZE** --W **SIZE**

💪 Citation

If you think this project is helpful in your research or for application, please feel free to leave a star⭐️ and cite our paper:

@misc{zhang2024vfimambavideoframeinterpolation,
      title={VFIMamba: Video Frame Interpolation with State Space Models}, 
      author={Guozhen Zhang and Chunxu Liu and Yutao Cui and Xiaotong Zhao and Kai Ma and Limin Wang},
      year={2024},
      eprint={2407.02315},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.02315}, 
}

💗 License and Acknowledgement

This project is released under the Apache 2.0 license. The codes are based on RIFE, EMA-VFI, MambaIR and SGM-VFI. Please also follow their licenses. Thanks for their awesome works.