Framer

Framer: Interactive Frame Interpolation

Wen Wang^1,2,   Qiuyu Wang²,   Kecheng Zheng²,   Hao Ouyang²,   Zhekai Chen¹,   Biao Gong²,   Hao Chen¹,
Yujun Shen²,   Chunhua Shen¹

¹Zhejiang University, China,   ²Ant Group

     

💡 Changelog

• [2024/12/20]. Thanks to kijai for the ComfyUI Support at ComfyUI-FramerWrapper!
• [2024/12/19]. Release the checkpoints.
• [2024/10/28]. Huggingface Gradio Demo is now available here!
• [2024/10/25]. Launch the project page and upload the arXiv preprint.

🔆 TL;DR

We propose Framer, a more controllable and interactive frame interpolation method that allows users to produce smoothly transitioning frames between two images by customizing the trajectory of selected keypoints, enhancing control and handling challenging cases.

Main Claims

The proposed method, Framer, provides interactive frame interpolation, allowing users to customize transitions by tailoring the trajectories of selected keypoints. This approach mitigates the ambiguity of image transformation, enabling much finer control of local motions and improving the model's ability to handle challenging cases (e.g., objects with differing shapes and styles). Framer also includes an "autopilot" mode that automatically estimates keypoints and refines trajectories, simplifying the process and enabling motion-natural and temporally coherent results.

Methodology

This work utilizes a large-scale pre-trained image-to-video diffusion model (Stable Video Diffusion) as the base model. It introduces additional end-frame conditioning to facilitate video interpolation and incorporates a point trajectory controlling branch for user interaction.

Key Results

Framer outperforms existing frame interpolation methods in terms of visual quality and natural motion, particularly in cases involving complex motions and significant appearance changes. Quantitative evaluation using FVD (Fréchet Video Distance) demonstrates superior performance compared to other methods. User studies show a strong preference for Framer's output, highlighting its effectiveness in producing realistic and visually appealing results.

Showcases

Note the videos are spatially compressed. We refer readers to the project page for the original videos.

1. Video Interpolation with User-Interaction

Start Image	Input Trajectory & Interpolation Results	End Image

2. Image Morphing with User-Interaction

Start Image	Input Trajectory & Interpolation Results	End Image

3. Video Interpolation without User-Input Control

Start Image	Interpolation Results	End Image

4. Novel View Synthesis

Start Image	Interpolation Results	End Image

5. Cartoon and Sketch Interpolation

Start Image	Interpolation Results	End Image

6. Time-lapsing Video Generation.

Start Image	Interpolation Results	End Image

Local Inference

Installation

conda create -n framer python=3.8 -y
conda activate framer
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.6 -c pytorch -c nvidia

pip install -r requirements.txt

Prepare weights

Download pretrained Stable Video Diffusion and Framer weights and put them in ./checkpoints, as follows.

checkpoints/
|─ stable-video-diffusion-img2vid-xt
└─ framer_512x320

Run gradio locally

python app.py

Some example inputs are presented in ./examples, enjoy the interactive frame interpolation!

📖 Citation BibTeX

Please consider citing our paper if our code is useful:

@article{wang2024framer,
  title={Framer: Interactive Video Interpolation},
  author={Wang, Wen and Wang, Qiuyu and Zheng, Kecheng and Ouyang, Hao and Chen, Zhekai and Gong, Biao and Chen, Hao and Shen, Yujun and Shen, Chunhua},
  journal={arXiv preprint https://arxiv.org/abs/2410.18978},
  year={2024}
}

🎫 License

For academic use, this project is licensed under the 2-clause BSD License. For commercial use, please contact C Shen.

😉 Acknowledgements

• Our implementation is based on SVD_Xtend and DragAnything.
• Thanks to kijai for the ComfyUI Support.