High Performance GMFSS with RIFE and GAN for Video Frame Interpolation
- Cupy is required as a running environment, please follow the link to install.
- The pre-trained model can be obtained from the following link, rename the folder to train_log and put it in the root directory, we provide two pre-trained models prototype_vanillagan and prototype_wgan.
python3 inference_video.py --img=demo/ --scale=1.0 --multi=2
2022.11.27, the training code has been made public. This code is run on a single V100, so there is no DistributedDataParallel related code. Please refer to RIFE train. The required training set is ATD-12K, which can be obtained from the official library. We also used a private data set, which unfortunately cannot be made public for the time being.In the process of organizing the code in the future, I found an error in the numerical processing of GAN Loss. I provided a replacement file, but the overall situation still maintained my original processing.