motion_blur_pytorch

Create smooth FPV drone footage, game recordings, or low-FPS videos using frame blending powered by PyTorch. Runs on Mac, Windows, and Linux.

Why?:

• repos like f0e/blur and smoothie can only run on Windows
• frame interpolation method used in those project can not interpolate large motions, and will produce artifacts

How?

• leveraging VFI advancements in recent years, we can improve the quality of frame interpolation by using the latest deep learning models like:
  • FILM
  • AMT
  • EMA-VFI
  • PerVFI
  • MoMo
  • and many more.

results

Before(24fps)	After(24fps)
1-before.mp4	1-after.mp4

insallation

• A fast GPU is required on linux and windows, for Mac you need to use M1~ M4 series chip with at least 32G RAM
• first, install pytorch follow the offical guide: guide
• install dependencies: pip install -r requirements.txt

AMT

download pretraind AMT-L model google drive or 百度网盘 password: iihx

edit file blur_amt.py

# change the path to your downloaded file
amt_model_path = "/Users/teli/www/ml/frame_interpolation/AMT/_pretrained/amt-l.pth"

process a folder with videos inside

python blur_amt.py -o output_folder_path input_folder_that_has_mp4_files/ 

process multi video files

python blur_amt.py -o output_folder_path path/to/a.mp4 path/to/b.mp4 path/to/c.mp4

FILM (a bit slow)

download pretraind FILM model google drive or 百度网盘 password: 8pxu

edit file blur_film.py

DEVICE = "mps" # cuda, mps
# change the path to your downloaded file
film_model_path = "/Users/teli/www/ml/frame_interpolation/frame-interpolation-pytorch-main/pretrained/film_fp16.pt"

process a folder with videos inside

python blur_film.py -o output_folder_path input_folder_that_has_mp4_files/ 

process multi video files

python blur_film.py -o output_folder_path path/to/a.mp4 path/to/b.mp4 path/to/c.mp4

TODO

• add more models

Limitations

• whithout quantization and other tricks, current implementation is slow
• requires a lot ram to run