New samplers

[Weatherby43]

Anyone have any info on the new samplers in the dev channel? (exponential and DPM++ 3M variants)? I've been looking for info on those, but so far haven't really found anything even talking about those... Just curious on best uses and what they are good at vs other samplers etc.

[catboxanon]

Exponential is the suggested scheduler for DPM-Solver++(2M) SDE (as opposed to Karras). See #12300 for more info. There is an option to override this for all k-diffusion samplers but it's now presented in the UI directly as an option for that particular sampler since it's recommended to be used with it, and makes specifying it as a separate sampler in hires fix and XYZ plots possible.

Regarding 3M: according to crowsonkb/k-diffusion#40 (comment), 3M requires more steps and also lower CFG. From Katherine's post in that thread (above the one I linked to) I'd gather it wasn't even added to k-diffusion previously due to this. 2M is a better choice since it doesn't have these limitations, but 3M is still being made available for those that want to try it out.

There were also more variations added a bit ago that change the solver to use Heun. There's a recent paper with more info in the relevant PR. #12542

---

Semi-related to your question, on the topic of schedulers: if you read anything about DDIM Uniform, LDM, and SGM: these are also just different schedulers. LDM and SGM just stand for "Latent Diffusion Models" and "Stability Generative Models" respectively, since those are the module names in the repos, and in the context of schedulers refer to the "vanilla" schedulers those repo implement. It would be pretty simple to add them to the webui if somebody wanted to.

[catboxanon]

Bumping this thread for more visibility now that 1.6.0 is released.