LLAVA is pretty great
As a "web builder" I do think these tools will be very useful for accessibility (eventually), specifically generating descriptive alt tags for images.
Locally it's actually quite easy to setup. I've made an app https://recurse.chat/ which supports Llava 1.6. It takes a zero-config approach so you can just start chatting and the app downloads the model for you.
What's your reservation about running it locally?
using it now on desktop (I am in China, so no OpenAI here) and in cloud cluster on project.