cut over from karras to model noise schedule for higher steps

[lstein]

The k_samplers come with a "karras" noise schedule which performs very well at low step counts but becomes noisy at higher ones.

This commit introduces a threshold (currently 30 steps) at which the k samplers will switch over from using karras to the older model noise schedule.

[psychedelicious]

Sounds like a good option to exposed for advanced users.

[db3000]

Is there a discontinuity there though, i.e. does 31 steps actually make it noisier than 30 steps because of the cutover?

[lstein]

Is there a discontinuity there though, i.e. does 31 steps actually make it noisier than 30 steps because of the cutover?

I've tested the region around the switchover and I can't see a discontinuity.

[lstein]

Sounds like a good option to exposed for advanced users.

I've added a "sticky" --karras_max option that allows you to adjust the switchover point. Can set to zero to use the LD noise model exclusively, or some high number (1000) to use the Karras schedule exclusively.

[jjoshua2]

Does newer, faster samplers like k_dpmpp_2 affect this 30 cutover? Because this one is so good that 10 is OK, 15 is normal, and 20 is high and 30 is never needed.