ComfyUI-Flux2Klein-Conditioning-Toolkit

Flux2KleinConditioningEnhancer

Use this when you want broad prompt-following adjustment on the active token region. Parameters: - magnitude: overall gain on active user tokens - contrast: sharpens or softens token-to-token differences while preserving average token energy - normalizestrength: equalizes token norms - preserveoriginal: blends some original conditioning back in, and in edit mode also reduces prompt pull against the reference path - preservemode: - blendafter: apply edits, then interpolate with original - dampen: reduce edit strength before applying edits - hybrid: dampen first, then blend back - edittextweight: extra gain when reference latents are present - activeendoverride: manual override when upstream metadata is incomplete - skipbos: default True; BOS should usually not dominate conditioning edits

entrada 1 · parâmetros 9 · saída 1

Flux2KleinPromptReferenceBalance

This is a convenience node. - balance = 0.0: mute prompt text, keep full reference - balance = 0.5: keep both - balance = 1.0: keep full prompt text, drive referencekeep to minimum Unlike the source repo, the reference side is implemented via reference mixing, not raw latent scaling. That means balance = 1.0 is only a true "reference off" endpoint when replacemode="zeros". With replacemode="lowpassreference", the weakened path still keeps a blurred coarse reference signal, including broad color / illumination structure and some channel statistics. Use that mode when you want to loosen detail lock without fully discarding appearance, not when you want the Prompt/Reference Balance slider to eliminate reference influence.

entrada 1 · parâmetros 6 · saída 1

Flux2KleinReferenceAppearanceBalancer

Use this when Klein keeps structure but washes out the source image's broad color/appearance cues during an edit. Parameters: - appearancescale: scales the blurred / low-frequency part of the reference latent - detailscale: scales the residual / high-frequency part of the reference latent - blurradius: controls how aggressively the latent is split into coarse appearance vs detail - channelmaskstart / channelmaskend: restrict effect to selected latent channels - targetreferenceindex: -1 = all references A practical starting point for washed-out edits is: - appearancescale = 1.15 to 1.35 - detailscale = 0.85 to 1.00 - blurradius = 2 to 4 This is intentionally more targeted than just increasing text contrast or globally scaling the whole reference latent.

entrada 1 · parâmetros 7 · saída 1

Flux2KleinReferenceLatentMixer

This node is intentionally framed as a mixer, not a “strength booster.” Parameters: - referencekeep: how much original reference latent to keep - replacemode: - zeros: strongest weakening / removal - gaussiannoise: loosens structure while keeping a noisy latent-like signal - channelmean: removes local spatial structure while keeping coarse channel statistics - lowpassreference: replaces removed content with a blurred copy of the original reference, which is often better for retaining broad color/illumination cues during edits - channelmaskstart / channelmaskend: restrict effect to latent channels - spatialfade / spatialfadestrength: fade only part of the spatial field - targetreferenceindex: -1 = all references

entrada 1 · parâmetros 8 · saída 1

Flux2KleinSectionedTextEncoder

- manual: use explicit front/mid/end inputs or [FRONT]...[MID]...[END] markers - autobalanced: split the prompt into rough 25/50/25 semantic sections by sentence/comma groups When the tokenizer exposes an attention mask, this node stores it in conditioning metadata so downstream nodes can operate on the real active token range.

entrada 1 · parâmetros 8 · saídas 5

Flux2KleinTokenRegionController

Use this when you want rough semantic steering by token position. It still uses positional regions, not semantic parsing, so the sectioned encoder is provided to make those positions more intentional.

entrada 1 · parâmetros 11 · saída 1