Paired video with lower resolution was used for training

Using ffmpeg can be generated downsampled video

ffmpeg -i input/mp4/video -s 960x540 -crf 30 -vsync vfr output/mp4/video

Model based on UNet architecture (more in tf_model.py)

Trained on MP4 encoded videos with lower resolution

Pretrained x2 TF model

Pretrained x3 TF model

foo@bar:~/VideoUpsampler$ python upsampler.py --file path/to/mp4/file --model path/to/model 

New video will be saved near original file with _upsampled suffix

Also suffix _audio will be added for video with copied audio