is there a tool to control strength of S_NOISE or ETA via mask

[AlexD81]

Hi, is there a tool to control strength of S_NOISE or ETA via mask/map? Need to reduce/increase S_NOISE for some parts of image. Of course, it can be done using mix of multiple ksamplers, but generation time become too high.

---

for those who interesting about NOISE and ETA, here's some info about them
https://sandner.art/sigma-and-eta-noise-image-variations-in-latent-space-and-fast-detail-effects/

[blepping]

@AlexD81 it's possible for s_noise (in a slightly roundabout way) with my SonarCompositeNoise node. the example actually demonstrates using stronger noise in some areas based on a mask: https://github.com/blepping/ComfyUI-sonar/blob/main/docs/advanced_noise_nodes.md#sonarcompositenoise

i don't think masks with ETA would make sense, ETA doesn't directly control the noise level. it's more like where the step begins and ends (and then based on ETA noise might get added to adjust the level to where a step normally would end). so to do masking with ETA you'd have to step to different sigmas in the actual sampler.

not exactly your question but if you want to have a lot of control over sampling you might also be interested in this project: https://github.com/blepping/comfyui_overly_complicated_sampling (OCS also has integration with ComfyUI-sonar, you can use all the custom noise types from Sonar with it.)

[ltdrdata]

FYI,
https://github.com/Extraltodeus/noise_latent_perlinpinpin
https://github.com/WASasquatch/PowerNoiseSuite
https://github.com/WASasquatch/PPF_Noise_ComfyUI

[blepping]

And grass often has "waves of gaussian noise".

that's probably not the case. :) you won't be able to see any kind of pattern in Gaussian noise. it's more likely your issue is related to the model and/or LoRAs you're using (possibly other parameters as well). however you might be able to use stuff like weird noise types to force the model out of its normal pattern. it's possible to chain the noise nodes (including the noise compositing nodes) from Sonar so you can mix however many noise types you want and use as many areas as you want (of course they may be a performance cost to generating and mixing many layers/types of noise).

the thing i've learned from my own generations/projects is that trying stuff and experimenting is rewarded. so my advice is to just try some different noise types at various strengths. you can accomplish a partial mix of weird noise + normal noise with something like this:

rescale rebalances the factors of the noise node and the ones connected to it to add up to the rescale value so this is 50% Gaussian noise, 50% Perlin noise. these noise types can be used with any sampler that adds noise during sampling (ancestral and SDE samplers) via the SamplerConfigOverride node. you can also use them as the initial noise for generation (or added noise in img2img) with NoisyLatentLike (also in Sonar). (rescale at 0 just disables rescaling, it doesn't scale everything to 0.)

FYI, https://github.com/Extraltodeus/noise_latent_perlinpinpin https://github.com/WASasquatch/PowerNoiseSuite https://github.com/WASasquatch/PPF_Noise_ComfyUI

on the subject of Perlin noise, you could possibly also try this branch of OCS: https://github.com/blepping/comfyui_overly_complicated_sampling/tree/perlin3d_experiment

it adds an advanced Perlin node that can generate a cube of 3D Perlin noise. each time noise is needed, it will take a 2D slice through the depth. it has an absurd number of tuneable parameters so you can get many different effects. (if you want to use it with stuff that takes SONAR_CUSTOM_NOISE there's also an adapter - not needed for OCS itself).

[AlexD81]

whank you for such detailed info! but I'm confused a bit with names. On this image

is rescale is a kind of "blend" value? and 1.0 is blend of 50%. Checked, it works, But what sould i set for, let's say, 30% ?
Tried rescale of 0.7 but result looks like reducing s_noise , so hard to understand how does it works

[blepping]

no problem. factor is the base strength of the generated noise. generally you want the base strength (which may get multiplied by other stuff later like s_noise, the sigma, etc but don't worry about that now) to be 1.0. one way to deal with when connecting multiple noise nodes is to manually set factor such that the factors of the connected nodes adds up to 1.0.

so in this example, you could set the two factors to 0.5 and you'd get a total of 1.0. however that can be difficult/time consuming if you have a bunch of noise nodes, which is where the rescale parameter comes in. first, if rescale is set to 0 that is considered disabled: it won't do any rescaling, you'll just get the factors added together.

when rescale is set to a non-zero value then the node with rescale enabled and the nodes connected to it will have their factors rebalanced to add up to the value in rescale. the rebalancing is proportionate so in this example the two 1.0 factor parameters would get rebalanced to 0.5. if you set factor in the Perlin noise node to 2.0 and left it as-is in the Gaussian node then the Perlin node's factor would get rebalanced to 2/3 and the Gaussian node's factor would get rebalanced to 1/3.

to put it a different less math way, if you connected several nodes like this

factor=1
   |
factor=1
   |
factor=1
   |
factor=2
   |
factor=1 rescale=1

you'd know they added up to the correct strength and the factor=1 nodes would have equal weight relative to each other and the factor=2 node would have double weight compared to the factor=1 nodes. the manual way to do it would be to disable rescale and set the factor=1 nodes to 1/6 and the factor=2 node to 1/3: (1 / 6) * 4 + 1 / 3 = 1.0

the way it works internally is summing the absolute value of the factors and then multiplying each node's factor by rescale / sum_factors.

But what sould i set for, let's say, 30% ? Tried rescale of 0.7 but result looks like reducing s_noise , so hard to understand how does it works

no need to use rescale for that and you're correct, setting rescale to 0.7 would be reducing the strength since the node and the ones it's connected to would have their factors rebalanced to add up to 0.7. rescale is mostly useful when you want to think about the noise type contribution proportionately, like "i want the Gaussian and Pyramid noise types to have equal strength and the Perlin noise type to have proportionately twice as much effect".

hope i didn't just end up confusing you more with this attempted explanation!

[AlexD81]

thank you for explanation, but it's appears too complicated especcially for my limited english knowledge. I'll back to this later, To be honest i'm still messing with my original question. if i need to increase/decrease noise in area specified by mask, what parameter should i change? . The problem is no matter what parameter i change, it also always affect unmasked area, inverse way. E.g increasing noise results to decreasing noise in unmasked area. Tried to enable normalize, but it didnt really help