Refactor of Stable Diffusion Handler in current PR

[w4ffl35]

Although many changes have been made to the sd_handler.py file over time, the core algorithm has remained the same. That is changing with this upcoming PR #590

I am finally refactoring the entire generate algorithm so as to reduce runtime and code. This is part of a larger effort to make images generate in near-realtime.

Why this feature?

I am running on older hardware - namely an RTX 2080s. As such my primary strategy when originally developing this application was to use a simple queuing system. I wasn't using threads in certain places that require it, and over all performance was a secondary concern.

When SD Turbo was released I wanted to test it and support it in AI Runner. When I did I saw immediate performance issues due to the architecture of the app and the speed of SD Turbo. As this revealed such large issues I decided to refactor with a concentration on performance.

Why "near" real-time?

As I mentioned I am on older hardware. I do not have the ability to test on anything about an RTX 2080s, so the fastest I can generate a 512x512 image at 20 steps is around 2 seconds (using SD 1.5). It is exponentially faster with SD turbo, but we're still not hitting anything close to 30fps or above.

Since I can't test on a faster card (and I don't know that ANY card can achieve real-time generation) I am calling this near-realtime. The aim is non-blocking, seamless image generation while drawing specifically.

I will post more updates here when I have this ironed out. I am aiming to complete this by tomorrow.

[w4ffl35]

We now have the ability to draw while images generate as well as interrupt the image generation automatically when the user begins to draw. This makes the experience of drawing to image generation more seamless and natural.

Take a look at the progression of this feature here: https://www.youtube.com/watch?v=m4AJQORHsew