A smaller, simpler, and faster version of https://github.com/qiuyu96/CoDeF
I built this to understand capabilities and limitations of the approach
cv::calcOpticalFlowFarneback
(rather than RAFT)Train
python3 run.py train --image_dir ./beauty_1
Train with a frame as the canonical image
python3 run.py train --image_dir ./beauty_1 --canonical ./beauty_1/00001.png
Generate frames
python3 run.py generate --checkpoint ./checkpoints/step=200.pt
Generate frames with a new canonical image
python3 run.py generate --checkpoint ./checkpoints/step=200.pt --canonical canonical.png