Nodejs + electron user interface for freyja subtitle OCR extractor.
Still in beta. All functionality are useable, but you may meet bug / crash when using the app. Please report any bugs
you meet with your log.log
file on github issues.
8GB of RAM required. Having a recent Nvidia GPU is strongly recommended or the process will be extremely slow.
-
If you are using Windows, please install Visual C++ Redist 2019. If you are using MacOS or Linux, make sure you have
ffmpeg
installed. -
Download latest version of Freyja from Releases page and extract it.
-
Download models from https://github.com/freyjaSubOCR/freyja-sub-ocr-model-zoo/releases. Download all txt and torchscript files, and place these files into
<app_path>/models/
folder. -
Run
freyja.exe
. EnableEnable CUDA
option if you have a Nvidia GPU, otherwise disable the option.
-
Video player is laggy.
Current video player implementation does not work well on real time video playback, a new implementation will be available when the app is out of beta.
-
MacOS and Linux versions do not work.
Currently there are some issues related with the underlying
torch-js
package. It should be fixed in the next beta version.
-
No audio on video playback.
Current video player cannot play audio.
-
Cannot play the video.
Maybe the video is an vfr (variable frame rate) video, which is not supported on current video player implementation. You can do a fast transcoding using ffmpeg to convert the video to a constant frame rate video:
ffmpeg -i video.mkv video_transcoded.mkv
. Remux won't work. -
Cannot use GPU models.
Make sure you have a recent Nvidia GPU. If you do have a Nvidia GPU, please try to update the driver.
-
The program says that "pyTorch backend crashed".
Please check the
log.log
.If the log says that
CUDA out of memory
, you need to reduce the batch size. If it still not works, it means that your GPU memory is too small and you can only use the CPU models.If the log shows other errors, please try to change the crop height of the video.