drhead

On another note, I do think this needs to include some sort of forced maximum interval between live preview updates as an option. Having it fully async like this is...

> I run a AMD rx 6700xt. I've swapped between this and what's in production, and noticed my speed goes down quite a bit using this commit. At 1024x with...

> It had nothing to do with my memory, same issues. In my own testing lately I have noticed a few issues with occasional noise outputs, but they're infrequent. I...

> This changes `single_sample_to_image` to return a tensor instead of PIL image; if some extension is using it, that extension will break. > > Is this only useful for grids,...

> It seems it interferes with #15751? That is a problem on their end, and can be resolved by not sending the sigmas to the device.

> I was testing with hr-fix and it seems it doesn't work? That's not a hi-res fix issue, it's a k-diffusion issue specific to that sampler (Euler and Heun at...

So, for this to resolve, either this PR needs to be merged upstream (https://github.com/crowsonkb/k-diffusion/pull/109) or I have to monkey patch the function. This probably isn't going to be a merge...

Now that the new schedulers are merged, I've went ahead and patched the to_d method there and have stripped the `device` argument from all schedulers as it is no longer...

> Is there a performance improvement? Yes. As stated, the current implementation induces a forced device sync during every sampling step, and the CPU will have to wait on the...

> What are the best conditions for observing the improvements? Doing 50-step 512x768 SD1 generations on 3090, and iterations per second between this and dev are not noticeably different. It's...