Important
This fork is focused on improving the user experience when inference video
Use batch_inference.py
to interpolate multiple videos at once.
Or check inference_video.py -h
for more options.
Note
Current model version: V4.17
# Install dependencies (notice that this will install torch with CUDA support, better run in venv or conda)
python3 -m pip install -r requirements.txt
# Interpolate a video
python3 inference_video.py /path/to/video.mp4
--ssim
: Program use SSIM to detect scene change, this argument can change the threshold of SSIM, default is0.4
, higher value means more sensitive to scene change, for a video with rapid change, you may want to lower this value (even set to0
) for less mistaking scene change detection.--scene_copy
: By default, if the scene change is detected, the model will stack the previous frame with the current frame to make the transition smoother (both 50% opacity). This argument will change this behavior to simply copy the previous frame as transition frame.--start_frame
: Start frame for inference, program will automatically detect unfinished work (stopped by hitP
key in preview window) and continue from there, so you dont need to specify this argument in normal cases.--start_time
and--stop_time
: Does same thing as--start_frame
, but with time in seconds. also useless in normal cases.--skip_frame
: Skip frame for inference, useful in case you want to re-interpolate a video with a new model version, but you have already deleted the original low-fps file, setskip_frame
to1
(or2
for 3x interpolation, etc) to just read the raw frames from a interpolated video.--resize
: Resize the input video to a specific scale, default is1.0
, set to0.5
for half size.--headless
: Run the program without preview window, useful for running on server or headless machine.