Cool cool, but hasn't this been achived by multiple projects many many times before?

I'd say that cost is a significant factor for doing this, and if you can leverage the usage of those fancy kubernetes GPU-cluster providers it'd come out even cheaper. Sure, it wouldn't be as "sophisticated", but it'd perform more or less identical and cost far far less.

All the above used cloud-providers are expensive, if compared to something like vast.ai or runpod.

Check this out; https://github.com/selkies-project/selkies