Extension for some Comfy/reForge samplers for Forge Neo (SEEDS, SA Solver, etc)

[Panchovix]

Hello guys, hoping you're doing well!

I did port as Haoming02 suggested, as an extension, some samplers that are on Comfy/reForge, but not natively on Forge Neo.

The link for the extension is: https://github.com/Panchovix/sd_forge_neo_extra_samplers

Added samplers for now:

• Gradient Estimation
• Gradient Estimation CFG++
• SEEDS 2
• SEEDS 3
• DPM++ SDE CFG++
• SA Solver
• SA Solver PECE

Any sampler can be added, but let me know if you want one in specific.

Let me know how it goes!

[Squibeel]

Can you maybe add the ones that are or work with RF?
You and Haoming have dif implementations of RF so its hard to reproduce the same/similar results on both webuis and idk if its because of the samplers.

[Panchovix]

Do you mean the RF samplers on reForge? I think Forge neo has euler RF/euler a RF and dpmpp 2s a RF (sample_dpmpp_2s_ancestral_RF), that are used automatically if you use a RF model instead of base euler/dpmpp 2s a.

Or do you mean other samplers in specific?

[Squibeel]

Since the results end up being so different between neo and reforge I wasn't sure if that was how it worked or not which is why I asked for them but if you say that that's how it works then I guess there's no need.

[Panchovix]

I will take a look nonetheless, but different results could be a different implementation as how RF itself works between the 2.

On Forge neo it is 1:1 to Comfy IIRC, while on reForge I had to do some weird patches to make it work.

[ArnorWing]

Hi. thanks for the extra samplers,
Any chance you can add HEUN_3S and HEUN_2S?
I've had great success with them and linear_quadratic.

[Panchovix]

Sure, will take a look today on evening/night.

EDIT: It seems those come from an extension? It seems the code is very convoluted, can't quite port it directly.

[ArnorWing]

I see you got HEUN2 in. thanks . I will test out now.
I have been using them in Comfy with the ClownSharkSampler and ZIT. Image quality is excellent but unable to do 4 separate seeds in a batch makes me want to keep using NEO

[zeydsama]

gbu💕💕

[elguachiiii]

anyway of adding scheduler with extention, i want ddim_uniform it gives very stable result with z-image compared to simple scheduler

[Panchovix]

@elguachiiii Should be possible, will take a look when I can.

[elguachiiii]

thanks i appreciate it

[anon0730]

As seemingly the only person in the world who uses DDPM (never seen anyone else mention it) I would like to request it. I swear it produces the cleanest results for SDXL after extensive sampler comparison and I settled on it when using ReForge so I would like it on Neo too.

[zeydsama]

do you still use sdxl nowdays?

[anon0730]

When I say SDXL I mean SDXL based models really (Illustrious and so on). Don't use them much now that Anima is out but some amount still and I would like to test DDPM on Anima too.

[raiken92]

Omg yes please! I've used DDPM almost exclusively for Illustrious/NAI models and it gives the best results. This was back when I was using reforge, but now that I switched to Neo I was really disappointed that there's no DDPM. So I also would like to request to have it on Neo as well..

[deprobot]

Thanks!

The one I'm missing most is DPM++ 2S a

[LastStardaughter]

Oooh. Has anyone made a list of which ones work with Anima?
I miss the "Karras Dynamic" scheduler.
I used to try the Euler Dy sampler sometimes but I don't actually know if it did anything better.