To be honest, all the heavy lifting is done by the depth-estimation model [1], otherwise this demo just wouldn't be possible.
I will warn you that viewing this may take some practice if you aren't already very well-versed in autostereogram viewing, and you will probably only want to run this on somewhat beefier laptop/desktop hardware, preferably in a Chromium browser.
For something less intense and mobile-friendly, I also made a sibling app that works with still images: https://namuol.github.io/magic-eye-anything
Both projects are open-source:
- https://github.com/namuol/magic-eye-mirror - https://github.com/namuol/magic-eye-anything
[1]: https://huggingface.co/spaces/Xenova/webgpu-realtime-depth-e...