DPM solver ++ third order with SDE

[christopher5106]

I see this in A1111... is it possible to get the missing implementation here: https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py#L887 Thanks (better results than 2nd order https://www.reddit.com/r/StableDiffusion/comments/16amqso/i_think_dpm_3m_sde_karras_is_producing_better/)

[yiyixuxu]

cc @asomoza, is this popular in the community? worth it to add this in?

[christopher5106]

Check the results here https://medium.com/@furkangozukara/sampling-method-dpm-2m-sde-vs-dpm-3m-sde-with-schedule-type-uniform-vs-karras-vs-exponential-09ab0ac379dd

[asomoza]

It's not that popular, almost all the good images I see don't use it, but there's some people that had claimed it's better.

I don't think the comparison done in that post it's really good, first, to really see if the result is good or not, you shouldn't use a lora and the same image and composition all the time. I get it that he needs to self promote but I don't think it's really scientific to put himself on all the images to do this.

My experience with it is not good so I don't use it anymore, there's a couple of issues with it that I don't like.

• It burns a lot faster and needs more steps or lower cfg.
• When you get a good result it's almost the same as the 2M

For example, what I like to use to compare them is to use plain color backgrounds over a subject, so we can see the colors and if it has artifacts.

I'll use comfyui and a couple of examples where I know it fails the most, both images use the same seed.

20 steps, 6.5 cfg, karras

dpmpp_2m_sde_gpu	dpmpp_3m_sde_gpu

Here the color spills more, the cat is more blueish in the 3M version.

20 steps, 8 cfg, karras	dpmpp_2m_sde_gpu	dpmpp_3m_sde_gpu

in this one the 3M adds a little more contrast and more details, it fails with the background since I asked a plain pink background.

30 steps, 8 cfg, karras	dpmpp_2m_sde_gpu	dpmpp_3m_sde_gpu

Here we can see that the 3M its burned, too much brightness and loses a lot of details. So to fix it we need to do a lot more steps.

50 steps, 8 cfg, karras	dpmpp_2m_sde_gpu	dpmpp_3m_sde_gpu

now it's fixed but almost the same as the 2M, and why bother if we get a good result with 2M at half the steps. You can also lower the CFG to fix it:

25 steps, 6 cfg, karras	dpmpp_2m_sde_gpu	dpmpp_3m_sde_gpu

Overall it's a lot more finicky and most of the time I don't really see the difference when they're both good, so I just stay with the 2M or other samplers.

I would also like the opinion of people that have good eyes like @bghira for this, since I don't consider myself that much of an expert evaluating samplers.

[christopher5106]

@asomoza can you confirm your tests are not under current regression https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/15519

[asomoza]

Oh I need to clarify that, I don't use automatic1111 anymore, I used comfyui for the examples. Also I don't really use SD 1.5 that much so I'm not qualified to test it with those models, because in my eyes I see them all bad. For example in that thread, people said that it lost "realistic textures" but honestly, I see all of them as plastic and fake.

I read some people that wrote that it's better with hands, so I'll test that later with both models, at least I can tell in SD 1.5 if the hands are bad.

Lastly I didn't test with anime because I'm not that good generating anime images and also I think that SD 1.5 and SDXL are mostly perfect at generating them with most samplers.

I encourage some more people to do some tests and post their findings here with high res images, don't use posts from other people and don't use LoRAs because the training of the LoRAs can make the results bad even if the sampler it's good.

[christopher5106]

Thanks :) very interesting points Originally I want this 3M SDE option for an upscaler with SD1.5.

[bghira]

i took a look and i'm wondering if there's a use case that isn't just still frames where this works better. if it works for upscaling, that's nice to have options, but the colour bleed seems substantial. has it been tested on eg. SV3D or SVD? controlnet commonly uses UniPC, maybe could work there too. but we're negating the benefits of UniPC by switching to a slower sampler if not

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.