Nodes/ComfyUI-Nunchaku-Klein/Nunchaku FLUX.2 Klein Enhancer
ComfyUI Node

Nunchaku FLUX.2 Klein Enhancer

Adjusts active FLUX.2 Klein conditioning embeddings.

By tom-m-2020·Created 26 days ago·Updated 6 days ago· 1
Nunchaku FLUX.2 Klein Enhancer
  • conditioning
  • CONDITIONING
active_scale1.00
per_token_whiten0.00
norm_equalize0.00
early_layer_scale1.00
mid_layer_scale1.00
late_layer_scale1.00
preserve_original0.00
active_end_override0
deviceauto
debugfalse
CategoryNunchaku/FLUX.2 Klein/Enhancer/Text

Inputs (11)

NameTypeDefaultDescription
conditioningCONDITIONING
active_scaleFLOAT1.000–10Multiplier on every active-token embedding. 1.0 = unchanged. The model was trained on Qwen3's natural distribution; values far from 1.0 push it off-distribution.
per_token_whitenFLOAT0.00-1–5Amplifies per-token deviation from the sequence mean: (x - mean)*(1+w) + mean. >0 widens spread, <0 compresses. Was called 'contrast' in v1.
norm_equalizeFLOAT0.000–1Blend each token toward the per-sequence mean L2 norm. Flattens magnitude variance — fights Qwen3's natural emphasis. 0 = no effect.
early_layer_scaleoptFLOAT1.000–5Klein-specific. Scale the first Qwen3 layer slice (low-level / structural features). Klein conditioning stacks 3 layers along the embed dim; this targets the first.
mid_layer_scaleoptFLOAT1.000–5Klein-specific. Scale the middle Qwen3 layer slice (intermediate semantic features).
late_layer_scaleoptFLOAT1.000–5Klein-specific. Scale the last Qwen3 layer slice (high-level / abstract semantic features).
preserve_originaloptFLOAT0.000–1Linear blend back the unmodified active region. 0.0 = full enhancement, 1.0 = no change.
active_end_overrideoptINT00–512Override the active-region end. 0 = auto-detect from attention_mask, falls back to full sequence length if mask missing.
deviceoptCOMBOauto2 options: auto, cpu
debugoptBOOLEANfalse

Outputs (1)

NameTypeDescription
CONDITIONINGCONDITIONING