Not a bug, just a question about this new quantization thing

[Goldenkoron]

Any more information about quantization? I am curious how it produces sharper and cleaner results and if it's good to use in all situations or not.

[C43H66N12O12S2]

#460

[ZeroCool22]

Any more information about quantization? I am curious how it produces sharper and cleaner results and if it's good to use in all situations or not.

Where you activate this settings?

Because i'm in settings tab but, i don't see it...

[Goldenkoron]

Any more information about quantization? I am curious how it produces sharper and cleaner results and if it's good to use in all situations or not.

Where you activate this settings?

Because i'm in settings tab but, i don't see it...

On the newest version,

[Birch-san]

you should use it in all situations. Stable Diffusion is trained on 1000 discrete timesteps/sigmas. you should only sample from the sigmas it was trained on. quantization fixes this for you. I think I saw code in the DDIM/PLMS samplers that already does quantization. but this wasn't the default for k-diffusion samplers.
I didn't find the difference to be dramatic. but it's slightly better, and the maths is more correct.
crowsonkb/k-diffusion#23