At 1 FPS this isn't quite real time enough to be useful for real time use cases, but I can see how a realtime stable diffusion would be a fun tool to use when producing for live events.

Combined with video interpolation techniques it may work: https://www.youtube.com/watch?v=sFN9dzw0qH8

Corridor Digital (YouTube channel) recently produced a video using almost exclusively img2img stable diffusion to redraw scenes into a cartoon/anime style.

I wonder if something like Redream could have sped up the production process rather than batch img2img.

That whole pipeline is kind of a performance nightmare... I think a much better approach would be raw diffusers, at the very least least.

I played with setting up a vapoursynth img2img pipeline myself (with VS doing the motion compensation to help with the temporal latent blending)

Is there a cpu-only version of this?

The base SD model run okay on CPU, and reasonably fast with intels MKL backend. But getting the whole auromatic1111 package to run on CPU seems to be a bit more complicated

Thought this was going to be the Dreamcast emulator

https://redream.io/

Love all these updates that are moving the open source scene forward, can't wait for a self-contained .exe install to try it.

Isn't the name "Redream" already the name of a commercial Dreamcast emulator?

My first thought as soon as I saw the title. It is the name of a Dreamcast emulator, however; as far as I know it is mainly free and offers some sort of cheap upgrade for certain features.

It's not open-source, that's for sure.

Commercial? Not really.

It's closed source freeware with a donation option.

What's the latest on temporal coherence?

IIRC there is a new controlnet model that is supposed to help.

But simple pre and post processing with latent blending can get you pretty far.