Future-proofing SD webui ?

[Ehplodor]

Stable Diffusion is only the latest step of open source generative models.

Given the peace of progress, I wonder about the future of SD models in general, as it is extremely probable to be outcompeted by the next generation of generative models, relatively soon.

When the time comes for us to use next-gen models, do we have to undergo all this work again, on the UI and scripts and all ? I mean, users and devs working hand in hand is wonderful, but time consuming. So maybe, if it is still possible, future-proofing all the hardwork that has been made here would be a good idea ?

[kungfooman]

Future proof code is code that is easy to read and extend and a common way to achieve this is usually:

1. typing
2. unit testing

The JS code can benefit from type checking through JSDoc (which also adds Intellisense + error hinting), just aswell as the Python code can benefit from typing.

And unit testing makes sure that new PR's don't create regressions, finding existing bugs, while also show-casing "how stuff is done".

[Ehplodor]

Also, modularity, and reusability ? Maybe via internal APIs.

Imagine SD v1.5, given it is a similar model to 1.4, "just more training", but same input and output. I believe it will be as some as cut and paste to replace v1.4.

However, what if v1.6 has more input and/or output parameters ? Or maybe 2.0 will be completely different architecture ? Then, will current SD repos have to be refactored / written from scratch again ?

[mayofiddler]

Might need to go to another language/interface for all that. Gradio is really just a prototyping tool. Maybe Fastai on top of Python and JupyterNotebooks, the way Fastai themselves run things? I saw somewhere that Google are going to start charging for commercial notebooks but you can run Jupyter locally as easily as SD. Just a thought, I have no axe to grind one way or another about any of it.

[phrozen]

Gradio supports API endpoints that I see are not working/documented. Maybe that is good place to start.