DPM2 always messes the image up on the very last step

[xvwyh]

I haven't seen this being mentioned anywhere else so I'm assuming this may be a problem that affects just me?

Disclaimer: I normally only use "Euler a" and "DPM2 a" samplers in practice, so I can only speak about my experience with these two, but for the sake of this post I tried others, and the same issue is present in LMS, Heun, and, to a lesser extent, PLMS. But here is my experience:

When using "DPM2" or "DPM2 a" samplers, the final few steps are always "enhancing" the image in some way, be it either adding more tiny details, or applying something that I can best describe as "photoshop filters" - sharpening, unsharp mask, playing with contrast etc. This is all fine and good, it makes the image better. But the VERY LAST STEP, regardless of how many steps it's configured to use, almost always makes the image very grungy/dirty and sort of stylized.

Images speak louder than words, so here are the final new steps saved with the "Save steps of the sampling process to files" script:

Example A (DPM2 a)

landscape
Steps: 20, Sampler: DPM2 a, CFG scale: 7, Seed: 3861236704, Size: 512x512, Model hash: 7460a6fa

Example A (DPM2)

landscape
Steps: 20, Sampler: DPM2, CFG scale: 7, Seed: 3861236704, Size: 512x512, Model hash: 7460a6fa

Example A (Heun)

landscape
Steps: 20, Sampler: Heun, CFG scale: 7, Seed: 3861236704, Size: 512x512, Model hash: 7460a6fa

Example A (LMS)

landscape
Steps: 20, Sampler: LMS, CFG scale: 7, Seed: 3861236704, Size: 512x512, Model hash: 7460a6fa

Example A (Euler a) - issue not present

landscape
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3861236704, Size: 512x512, Model hash: 7460a6fa

Example A (Euler) - issue not present

landscape
Steps: 20, Sampler: Euler, CFG scale: 7, Seed: 3861236704, Size: 512x512, Model hash: 7460a6fa

Example B (DPM2 a)

photo of a young woman, instagram
Steps: 20, Sampler: DPM2 a, CFG scale: 7, Seed: 3924002472, Size: 512x512, Model hash: 7460a6fa

I'm quite happy with generation results of "DPM2 a", so I'd like to continue using it, but because of this I'm forced to always manually interrupt the generation before the final step, otherwise the result will be, well, ruined, and I'd have to either manually save the pre-final step (if I used the aforementioned script), or generate the image again and interrupt it in time.

Is there any way to avoid this? Automate the skip of the very last step? Or fix whatever this issue of mine is that for some reason only affects the last step?

[Ehplodor]

Seems already corrected in PR#1560 (sorry, I don't know how to reference a PR from smartphone)

[xvwyh]

The script from #1560 indeed does the trick, assuming it doesn't negatively impact the generation in some other way, which cannot easily be tested because it breaks seed compatibility, as mentioned by the author. Thank you.