[Feature]: Fixing of Sampler Flaws (falsely interpolated steps / limiting results to medium brightness)

[TheOnlyHolyMoly]

Feature description

Reminder for conversation (#1091 (reply in thread)) as requested by @vladmandic

Mainly the issue narrows down to issues with the timesteps function leading to a Non-Zero Terminal Noise Ratio, leading to a bias in at least DDIM samplers, but possibly other samplers that follow the timesteps approach.

Details and solution approach captured in above linked discussion thread.

Resources
huggingface/diffusers#2585
DDIM Repo Owner >https://github.com/ermongroup/ddim
Scientific Paper on problems & fixing: https://arxiv.org/abs/2305.08891

Version Platform Description

Affects all platforms

[DirtyHamster]

There was a live stream going over this paper last week: https://www.youtube.com/watch?v=o1eJAUe-czQ

[TheOnlyHolyMoly]

Adding in. Looks like diffusers team is working on the same matter.

huggingface/diffusers#3664

[vladmandic]

implemented in diffusers, no plan to bring it to original backend at the moment.