Using with start/end step ksamplers (such as in refiner workflows)?

[Jonseed]

Some have asked if Detail Daemon or Lying Sigmas can be used in workflows with the advanced ksamplers, which have start and end steps values, which allow for refiners and such. I'm not sure it will work with those core nodes since they have the samplers built-in. It would require an advanced ksampler with the option to input a custom sampler. So far I haven't found a suitable custom node. Do you know of any @blepping ?

[blepping]

It would require an advanced ksampler with the option to input a custom sampler. So far I haven't found a suitable custom node.

yeah, unfortunately you basically lose the ability to input parameters with the non-custom samplers.

i probably don't know something that does what you're looking for, but also the question was pretty general so i'm not sure exactly what feature the user is looking for. there is one possible workaround (though it definitely has limitations). you can add samplers to the list users can choose from so you could do something like add a euler_detail_daemon (or whatever) with some reasonable preset values. that way people could choose it from the list, though they wouldn't be able to change parameters themselves.

i'd use that approach sparingly to avoid cluttering up the list with samplers only some people would actually want to use though, so probably keep it to 1-2 presets at most (unless you add configuration or something that lets the user specifically opt into it). also would be possible to do something like let the user create a config file in the custom node directory like some nodes do which allows defining some presets that can be added to the list when the node loads.

[Jonseed]

@blepping here is what they were looking for:

"I wanted to know if its possible to use Detail daemon with the default Ksampler (Advanced), because I wanted to use "start at" and "end at" steps and "total steps" options. It seems that SamplerCustom doesn't have these options. I use two different ksamplers to control the step at which generation stops and then send noise to the other one, it's easier this way to control which lora to load and unload."

I haven't used that particular kind of workflow before with LoRAs, but I know SDXL uses a similar base/refiner workflow.

Impact Pack has a KSamplerAdvancedProvider node that looks like it lets you specify an input sampler, but I'm not sure this would work with the traditional KSamplerAdvanced with start and end steps...

[blepping]

because I wanted to use "start at" and "end at" steps and "total steps" options.

you can do that by slicing up the sigmas with a SplitSigmas node.

i believe this would be the equivalent of start at step 2, end at step 4.

high_sigmas is the sigmas before the split, low_sigmas is after.

edit: might be necessary to subtract one from the steps since SplitSigmas is zero based. not sure if the advanced sampler is or not. i.e. step 0 is the first step, step 1 is the second, etc.

[bluewanderer]

God! That's so brilliant!

[Jonseed]

I found a RegionalSamplerAdvanced that accepts a custom sampler, and has start and end steps. I wonder if this workflow could work.

[blepping]

actually, thinking about this some more (and running into my own situation where i wanted to use it but couldn't) i guess there is another alternative - make a model patch version. the user would need to pass it sigmas that match whatever the sampler is using relatively closely (probably doesn't have to be exact because of the interpolation). i think they'd probably also have to manually set CFG.

ComfyUI's normal model patching stuff only works for SD1x/SD2x/SDXL though so at least the first pass would only support those models, no Flux or SD3x.

[Tastypotato0009]

Thank you! I followed your instructions, and now DD works with KSampler using BlehSetSamplerPreset!

[Jonseed]

Maybe I need to add some notes to the README about how to use DD and LyingSigmas with non-custom samplers, or at least point them to this discussion.

[Jonseed]

@blepping will the Multiply Sigmas node also work with the BlehSetSamplerPreset node, inputting the adjusted sigmas into the override_sigmas_opt?

[blepping]

will the Multiply Sigmas node also work with the BlehSetSamplerPreset node, inputting the adjusted sigmas into the override_sigmas_opt?

it will work as in the feature will function as intended but it might not do what you want. those override sigmas only apply to what gets passed to the sampler (not a node that does sampling like a KSampler, sampler as in Euler). here's an example of how you might run into unintended effects using multiplied sigmas (or different schedules/denoise levels) in the sigma override:

suppose you generate some sigmas and pass them to the preset sigma override: 1.1, 0.6, 0.0

then you either pass sigmas to the KSampler (or whatever) or it generates them internally: 1.0, 0.5, 0.0

The KSampler will add initial noise based on sigma 1.0, but you're actually sampling from sigma 1.1 since that's what got passed to the sampler so not enough noise will be added and you'll get less detail.

so the important thing to be aware of is if you override the sigmas like that, if the sampler node is starting from a different sigma there's probably going to be a mismatch between the amount of noise you're sampling with and the amount the sampler added. this is definitely not always bad: i often use a multiplier for the initial noise - just something to be aware of.

one way around it (which unfortunately doesn't work for RF models like Flux/SD3x) is to disable adding noise in the sampler and add it yourself using something like my NoisyLatentLike node from ComfyUI-sonar:

if you disable add noise, watch out for this bug which causes non-deterministic generations: comfyanonymous/ComfyUI#4518 (bleh has a workaround for this which could be combined with the sampler preset - BlehForceSeedSampler will ensure the seed always gets set before sampling starts.)

[freshwind-bit]

Many thanks @Jonseed for the Detail Daemon Sampler and @blepping for such an elegant solution to integrate this node into a workflow with a variable step range of the chain of sequential advanced samplers (like in Dr. Lt. Data's video example with skipping one generation step and doubling the next one). Works great and is very effective both in terms of increasing and decreasing detail while maintaining the overall composition. Drastically game changer!