Nodes/ComfyUI-DazzleKSampler/Dazzle KSampler (DazzleNodes)
ComfyUI Node

Dazzle KSampler (DazzleNodes)

A ComfyUI node in RES4LYF/samplers with 22 inputs and 3 outputs.

By DazzleNodes·Created 5 months ago·Updated 2 months ago· 3
Dazzle KSampler (DazzleNodes)
  • model
  • positive
  • negative
  • latent_image
  • sigmas
  • guides
  • options
  • output
  • denoised
  • options
eta0.50
sampler_nameres_2m
schedulerbeta57
steps30
steps_to_run-1
denoise1.00
cfg5.50
seed0
latent_roleauto
sampler_modestandard
bongmathtrue
noise_allgaussian
noise_type_initgaussian
noise_type_sdegaussian
noise_type_sde_substepgaussian
CategoryRES4LYF/samplers

Inputs (22)

NameTypeDefaultDescription
etaFLOAT0.50-100–100Calculated noise amount to be added, then removed, after each step.
sampler_nameCOMBOres_2m119 options: none, res_2m, res_3m, dpmpp_2m, dpmpp_3m, abnorsett_2m, +113
schedulerCOMBObeta5710 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +4
stepsINT301–10000
steps_to_runINT-1-1–10000
denoiseFLOAT1.00-10000–10000
cfgFLOAT5.50-100–100
seedINT0-2–18446744073709550000
latent_roleCOMBOautoHow to interpret the input latent dict. 'auto' inspects dict shape (use_as_noise flag, presence of 'noise' key) and dispatches accordingly. 'noise' = use upstream's noise tensor (samples for Shape B, noise key for Shape C); init zeroed. 'latent_image' = use samples as init image, generate noise from seed. 'noise+latent_image' = use 'noise' key as noise + samples as init (Shape C only; falls back otherwise). 'seed_driven' = true txt2img: zero init AND generate noise from seed (distinct from latent_image, which preserves samples as img2img init). Mismatch warnings printed to console when explicit role does not match incoming dict shape. seed=-2 magic is deprecated under 'auto'. Note: latent_role controls only the initial noise tensor at sigma_max -- the seed still drives per-step ancestral/SDE noise injection regardless of role. Set eta=0 for output fully determined by upstream noise.
sampler_modeCOMBOstandard3 options: unsample, standard, resample
bongmathBOOLEANtrue
noise_allCOMBOgaussianMaster noise selector. When set to anything other than 'custom', overrides all three noise_type_* widgets below with this value (the sub-widgets are auto-hidden by the JS extension). Set to 'custom' to expose noise_type_init, noise_type_sde, and noise_type_sde_substep for independent configuration. DazNoise:* options appear when dazzle-comfy-plasma-fast is installed.
noise_type_initoptCOMBOgaussianAdvanced. Leave at gaussian for default behavior. Spectrum of the *initial* noise tensor at sigma_max (only consumed when latent_role generates noise from seed -- ignored when latent_role=noise/noise+latent_image because initial noise then comes from upstream). Brown/pink emphasize low frequencies (smoother/painterly), blue/violet emphasize high frequencies (grainier/sharper), plasma/pyramid produce fractal/structured patterns. DazNoise:* options (Plasma/Pink/Brown/Greyscale/Gaussian) appear when dazzle-comfy-plasma-fast is installed -- they shape the initial canvas the model imagines into; for clean composition use SmartResCalc upstream with spectral_blend instead, since these are raw structured noise.
noise_type_sdeoptCOMBOgaussianAdvanced. Leave at gaussian for default behavior. Spectrum of the *per-step* ancestral/SDE noise injected at every sampler step (scaled by eta). Active for ALL latent_role values -- this is the dominant compositional noise budget under latent_role=noise. Same spectral choices as noise_type_init. Caveat: highly structured noise types (raw plasma, brown, pyramid) produce visual artifacts (chromatic fringing, halftoning) when used as per-step injection because the diffusion model expects ~unit-variance Gaussian statistics; for shaped per-step noise without artifacts, use SmartResCalc upstream with spectral_blend and latent_role=noise. DazNoise:Gaussian is safe (true Gaussian).
noise_type_sde_substepoptCOMBOgaussianAdvanced. Leave at gaussian for default behavior. Spectrum of per-substep noise injection (only relevant for samplers that use multi-substep SDE, e.g. some RES4LYF RK variants). Same caveat as noise_type_sde for highly structured types.
modeloptMODEL
positiveoptCONDITIONING
negativeoptCONDITIONING
latent_imageoptLATENT
sigmasoptSIGMAS
guidesoptGUIDES
optionsoptOPTIONS

Outputs (3)

NameTypeDescription
outputLATENT
denoisedLATENT
optionsOPTIONS