The decoder is written in Python, tested against Python 3.10. It uses some newer features from the typing module so older versions probably won't work.
The only external dependencies required are numpy and Pillow.
python3.10 -i decode.py subject.jpg will drop you into a shell where you can play with the PIL image and the intermediate metadata structures. Try image_segments to get a list of all segments, or image_metadata to get the consolidated, parsed metadata structure. You can also look at decode_animation.py as an example of what kind of things you can do using the parsed results.
python3.10 decode.py subject.jpg output.png will write the PIL image to a PNG file.
Note: random JPG files you find on the internet probably won't work with this decoder. The feature set is very small, and it still has a lot of bugs. It's talk-ware, for sure :)
$ npx @marp-team/marp-cli@latest slides.md