[FFmpeg-user] Training or finding pretrained models for the sr video filter

[Vítek Novotný]

Dear everyone,

I am finding myself increasingly using the DNN superscalers for textual
and drawn images and videos. Thus far, I was content using the waifu2x
image superscaler [1]. However, when processing videos, exploding them
to frame images is space-consuming and superscaling them one by one is
time-consuming. To eliminate the space consumption concern, I was
looking for an ffmpeg video filter, which would save me the trip from
video to frame images and back.

In the ffmpeg documentation, I discovered the existence of the sr video
filter [2], which claims to be able to do just this. However, there seem
to be no pre-trained DNN models in the filter repository [3] and the
documentation for producing models is severely lacking: Dependencies for
running the Python scripts are underspecified and the training scripts
keep failing in random places due to missing Python packages and wrong
version of Python. Is there any better documentation or, better yet,
some pretrained models to get started?

 [1]: https://github.com/nagadomi/waifu2x
 [2]: https://ffmpeg.org/ffmpeg-filters.html#sr-1
 [3]: https://github.com/XueweiMeng/sr

Kind Regards,
Vit Novotny
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <https://ffmpeg.org/pipermail/ffmpeg-user/attachments/20200417/081cc67b/attachment.sig>