Spacevid

Spacevid is a tool for transcoding large videos to fit a specified file size with minimal adjustment to the bitrate.

Limitations

This application was developed specifically for my use case so it only outputs 480p 24FPS WebMs, the only thing that varies is the output bitrate.

Prerequisites

• FFMPEG
• Tensorflow
• Python3

Usage

Dump a bunch of videos in ./dataset/videos then run ./generate_data.py, for best results you'll want to use videos similar to the ones you're likely to convert (e.g: if you tend to convert alot of YouTube and TikTok clips, use those kinds of videos).

After you've generated a bunch of data (give it a few hours or so) you can use ./convert.py to convert your video. It takes the following arguments:

• filename - The path to the video file you want to convert
• size - The target file size in bytes
• -o --output - The path you'd like to output to
• -e --epochs - The amount of training sessions you'd Spacevid to do before converting
• -d --dont-remember - By default Spacevid will save the results of every transcode and use it to train itself in future, use this if you'd like to avoid that
• -h --help - show the help (it's basically just this tbh)

Background

I wrote it while on holiday because I had almost no internet and so I started trying to transcode a bunch of video memes I had so that they'd be small enough to post on Spacechan. The formulas I found would give me bitrates that produced files that were way too small and low quality and I didn't want to figure it out through trial and error. It then occured to me that a neural net would be perfect for this and it'd give me an opportunity to learn Tensorflow so that's what I did.
Thanks to ∆xel for giving me pointers.