Some days ago, OpenAI released the code and models for Shap-E, which lets you do text-to-3D, and download a mesh you can use in any 3D software or for game development (rather than NeRFs with extra steps like previous models, and many papers that never released their code). This is very exciting, as the quality is reasonably good, and previously I would try to get that through various hacks.
There is already a HuggingFace space to try it, but no model on HuggingFace that you can easily use with their Inference API. You can fork the space and turn it into an API with some extra work, but I wasn't able to easily figure this out, and since running your own space is expensive anyway, I decided to take the easy way out and put a Flask server on top of OpenAI's existing code.
My server implementation is here. Since generating a new model only takes a couple seconds, I decided to design the interface as a "just-in-time" download link. You request /models/cat.ply, and if it's been generated before, it will download right away, but if not, then it's generated on the fly while the request is blocking.
I ran this on vast.ai, on an A10 instance, but I'm probably not going to keep it up for long as it's a bit expensive. I used the default pytorch image, and tweaked the config to expose the server on port 8081, which is a Docker port that vast.ai then maps to a port on the host. I added a proxy to that on my Model Compositor project which you can try here for free.