RIFE - Real Time Video Interpolation

This project is the implement of RIFE: Real-Time Intermediate Flow Estimation for Video Frame Interpolation. If you are a developer, welcome to follow Practical-RIFE, which aims to make RIFE more practical for users by adding various features and design new models.

Currently, our model can run 30+FPS for 2X 720p interpolation on a 2080Ti GPU. It supports 2X,4X,8X... interpolation, and multi-frame interpolation between a pair of images.

16X interpolation results from two input images:



CLI Usage


git clone [email protected]:hzwer/arXiv2020-RIFE.git
cd arXiv2020-RIFE
pip3 install -r requirements.txt


Video Frame Interpolation

You can use our demo video or your own video.

python3 inference_video.py --exp=1 --video=video.mp4 

(generate video_2X_xxfps.mp4)

python3 inference_video.py --exp=2 --video=video.mp4

(for 4X interpolation)

python3 inference_video.py --exp=1 --video=video.mp4 --scale=0.5

(If your video has very high resolution such as 4K, we recommend set --scale=0.5 (default 1.0). If you generate disordered pattern on your videos, try set --scale=2.0. This parameter control the process resolution for optical flow model.)

python3 inference_video.py --exp=2 --img=input/

(to read video from pngs, like input/0.png ... input/612.png, ensure that the png names are numbers)

python3 inference_video.py --exp=2 --video=video.mp4 --fps=60

(add slomo effect, the audio will be removed)

python3 inference_video.py --video=video.mp4 --montage --png

(if you want to montage the origin video, skip static frames and save the png format output)

The warning info, 'Warning: Your video has *** static frames, it may change the duration of the generated video.' means that your video has changed the frame rate by adding static frames, it is common if you have processed 25FPS video to 30FPS.

Image Interpolation

python3 inference_img.py --img img0.png img1.png --exp=4

(2^4=16X interpolation results)
After that, you can use pngs to generate mp4:

ffmpeg -r 10 -f image2 -i output/img%d.png -s 448x256 -c:v libx264 -pix_fmt yuv420p output/slomo.mp4 -q:v 0 -q:a 0

You can also use pngs to generate gif:

ffmpeg -r 10 -f image2 -i output/img%d.png -s 448x256 -vf "split[s0][s1];[s0]palettegen=stats_mode=single[p];[s1][p]paletteuse=new=1" output/slomo.gif

Run in docker

Place the pre-trained models in train_log/\*.pkl (as above)

Building the container:

docker build -t rife -f docker/Dockerfile .

Running the container:

docker run --rm -it -v $PWD:/host rife:latest inference_video --exp=1 --video=untitled.mp4 --output=untitled_rife.mp4
docker run --rm -it -v $PWD:/host rife:latest inference_img --img img0.png img1.png --exp=4

Using gpu acceleration (requires proper gpu drivers for docker):

docker run --rm -it --gpus all -v /dev/dri:/dev/dri -v $PWD:/host rife:latest inference_video --exp=1 --video=untitled.mp4 --output=untitled_rife.mp4


Download RIFE model or RIFE-Large model reported by our paper.

UCF101: Download UCF101 dataset at ./UCF101/ucf101_interp_ours/

Vimeo90K: Download Vimeo90K dataset at ./vimeo_interp_test

MiddleBury: Download MiddleBury OTHER dataset at ./other-data and ./other-gt-interp

HD: Download HD dataset at ./HD_dataset. We also provide a google drive download link.

python3 benchmark/UCF101.py
# "PSNR: 35.246 SSIM: 0.9691"
python3 benchmark/Vimeo90K.py
# "PSNR: 35.506 SSIM: 0.9779"
python3 benchmark/MiddleBury_Other.py
# "IE: 1.962"
python3 benchmark/HD.py
# "PSNR: 31.99"
python3 benchmark/HD_multi.py
# "PSNR: 18.89(544*1280), 28.83(720p), 24.96(1080p)"

Training and Reproduction

Because Vimeo90K dataset and the corresponding optical flow labels are too large, we cannot provide a complete dataset download link. We provide you with a subset containing 100 samples for testing the pipeline. Please unzip it at ./dataset

Each sample includes images (I0 I1 Imid : 9 x 256 x 448), and optical flow (flow_t0, flow_t1: 4, 256, 448).

For origin images, you can download them from Vimeo90K dataset.

For generating optical flow labels, our paper use pytorch-liteflownet. Please notice that due to the data augmentation during training, generating optical flow labels during training may cause performance loss. We recommend that readers use RAFT to generate optical flow labels because it is easier to deploy. As long as the generated labels are correct, our method is not sensitive to the teacher model. Because the teacher can see the intermediate frame, the quality of the generated optical flow is much higher than that of the student.

We use 16 CPUs, 4 GPUs and 20G memory for training:

python3 -m torch.distributed.launch --nproc_per_node=4 train.py --world_size=4


  title={RIFE: Real-Time Intermediate Flow Estimation for Video Frame Interpolation},
  author={Huang, Zhewei and Zhang, Tianyuan and Heng, Wen and Shi, Boxin and Zhou, Shuchang},
  journal={arXiv preprint arXiv:2011.06294},