Agreed with the artist part. My previous startup (headlime.com, acquired by Jasper.ai), was mostly used by copywriters instead of non savvy folks. I genuinely believe that text2image AI will help artists, not hurt them.
I have an open source example using openai https://observablehq.com/@tomlarkworthy/tarot-backend
You should try to host the model yourself.
Using the SD API is great for a POC, but to scale this and take full advantage of the cost benefits, you need to spin up a cloud GPU.
Heck - you can even build an off the shelf PC and pull workloads from the cloud, and then upload the resulting images. That would work well within your traffic needs. Maybe the upfront cost isn't worth it to you, but if you already have the components ...
(I've done this myself. I run a dual cloud / on-prem 3090 cluster.)
It's just a task library (I used Dramatiq) with a worker that connects to it. Then the worker picks up tasks, processes them and uploads the results to R2.