Comment by pseudosavant
4 days ago
I've been using FFMPEG for 15+ years, and still can't remember almost any commands. LLMs have been amazing for using FFMPEG though. ChatGPT and Claude do wonders with "give me an ffmpeg command that will remux a video into mkv, include subtitle.srt in the file, and I only want it between 0:00:05 and 0:01:00." It produced this in case you were wondering: `ffmpeg -i input.mp4 -i subtitle.srt -ss 00:00:05 -to 00:01:00 -map 0 -map 1 -c copy -c:s mov_text output.mkv`
I wonder how small of an LLM you could develop if you only wanted to target creating ffmpeg commands. Perhaps it could be small enough to be hosted on a static webpage where it is run locally?
Perhaps small enough to include in ffmpeg itself so you can just write commands `ffmpeg do this thing I want`.
Now I say this, it seems like there should already be a shell that is also an LLM where you can mix bits of commands you vaguely remember and natural language a bit like Del Boy speaking French...
That would be an amazingly useful feature of ffmpeg, and considering how large its dependencies are (390MB of packages for `apt install ffmpeg` on a fresh Raspberry Pi OS install), it would be reasonable to have an optional model package.
Warp terminal does that. It's cool.
https://github.com/timschmidt/llmterm
-c:s mov_text is unnecessary and in fact might be fucking things up
I was kind of curious about that one too. It is encoding the SRT into the MP4/MOV subtitle format. I use it all the time when muxing subs into MP4s, but I haven't seen what happens with an MKV like that. It is very well supported in MP4s.