This is a video frame interpolation algorithm implemented using VideoFlow, which uses multiple frame inputs and ranked first in Sintel Clean.
2023-10-04: We have updated an anime model and trained it for 60 epochs on our proprietary dataset, and we will soon release a new model that can better gather spatiotemporal information.
git clone https://github.com/hyw-dev/VFSS.git
cd VFSS
pip3 install -r requirements.txt
- Download the pretrained models from here.
- Move the 'anime' and 'real' folders to train_log/*
You can use demo video or your own video.
python3 inference_video.py --exp=1 --video=video.mp4 --model_type [anime/real]
(generate video_2X_xxfps.mp4)
python3 inference_video.py --exp=2 --video=video.mp4 --model_type [anime/real]
(for 4X interpolation)
python3 inference_video.py --exp=1 --video=video.mp4 --scale=0.5 --model_type [anime/real]
(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/ --model_type [anime/real]
(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 --model_type [anime/real]
(add slomo effect, the audio will be removed)
python3 inference_video.py --video=video.mp4 --montage --png --model_type [anime/real]
(if you want to montage the origin video and save the png format output)
Download Vimeo90K dataset(septuplet).
We use 1xV100 for training (Note that our real model only trained 12 epoch on the vimeo septuplet):
python train.py
Optical Flow: VideoFlow
Video Interpolation: GMFSS RIFE softsplat
This project is sponsored by SVFI Development Team