Nodes/KreaSeedVarianceEnhancer/Seed Variance Enhancer - Krea 2 Turbo
ComfyUI Node

Seed Variance Enhancer - Krea 2 Turbo

A ComfyUI node in advanced/conditioning with 12 inputs and 2 outputs.

By harukimix·Created about a month ago·Updated about a month ago· 14
Seed Variance Enhancer - Krea 2 Turbo
  • conditioning
  • conditioning
  • diagnostics
randomize_percent50
auto_strength_factor1.00
strength20.0000
noise_insert
steps_switchover_percent25
seed0
mask_starts_at
mask_percent0
log_to_consolefalse
noise_distribution
granularity
Categoryadvanced/conditioning

Inputs (12)

NameTypeDefaultDescription
conditioningCONDITIONING
randomize_percentFLOAT501–100Percentage of embedding values (or tokens, see granularity) that receive noise. Higher = more diversity, less prompt adherence.
auto_strength_factorFLOAT1.000–100If > 0, noise scale = measured embedding std * factor (recommended for Krea 2 Turbo). Try 0.5 subtle / 1.0 balanced / 3+ strong. Set to 0 to use the manual 'strength' field instead.
strengthFLOAT20.0000-4294967295–4294967295Absolute noise scale. Only used when auto_strength_factor is 0.
noise_insertCOMBOWhere the noisy embedding is used. 'beginning steps' perturbs the structure-deciding early steps, then restores the clean prompt. 'all steps' is strongest but worst for adherence/text.
steps_switchover_percentFLOAT251–99Where noisy/clean conditioning switch. Formula: (100/TOTAL_STEPS)*SWITCH_STEP. For 8-step Krea 2 Turbo: 25 ~ first 2 steps, 38 ~ first 3 steps.
seedINT00–18446744073709550000IMPORTANT: set this to 'randomize' or 'increment' to get DIFFERENT variations each run. If left 'fixed', every generation gets the SAME embedding noise and outputs stay similar.
mask_starts_atCOMBOWhich end of the prompt is protected from noise.
mask_percentFLOAT00–99Percentage of the prompt protected from noise. 'beginning'+'50' protects the first half.
log_to_consoleBOOLEANfalseAlso print the diagnostics to the ComfyUI console. The same text is always available on the 'diagnostics' STRING output. Version k2-1.2.0 (base: SVE 2.2).
noise_distributionCOMBOShape of the random noise. 'uniform' is the original behavior. 'gaussian' matches the bell-curve statistics of the embedding itself and tends to produce fewer broken outputs at the same strength. Both are power-matched: the same strength gives the same overall perturbation.
granularityCOMBO'values' scatters noise across individual embedding values (original behavior). 'tokens' perturbs whole tokens at once, producing more semantic variations (subjects/props/composition shift together). randomize_percent then selects tokens instead of values.

Outputs (2)

NameTypeDescription
conditioningCONDITIONING
diagnosticsSTRING