DDRK Flux Conditioning

Inspect and reshape text conditioning: padding attenuation, token gain and norm equalisation. Prints what it did.

By HVOSTOVSKY·Created 2 months ago·Updated about 17 hours ago· 0

conditioning

CONDITIONING

◄pad_attenuation0.00►

◄active_gain1.00►

◄norm_equalize0.00►

◄front_scale1.00►

◄mid_scale1.00►

◄end_scale1.00►

◄reporttrue►

Categorysampling/DDRK Omega/utils

Inputs (8)

Name	Type	Default	Description
conditioning	CONDITIONING		—
pad_attenuation	FLOAT	0.000–1	Scales padding tokens toward zero. 0 = untouched (stock Flux), 1 = fully zeroed (what Chroma does). Only acts when the encoder supplied an attention_mask; without one the padding boundary is unknown and this is skipped. Try 0.5-1.0 on short prompts, where padding dominates the sequence.
active_gain	FLOAT	1.000.1–3	Multiplier on real (non-padding) tokens. 1.0 = off. Raising it strengthens prompt influence in a blunt way; large values distort the embedding geometry the model was trained on, so stay near 1.
norm_equalize	FLOAT	0.000–1	Blends per-token vector norms toward the mean norm of the active region. 0 = off, 1 = every active token has equal norm. Attention weight scales with token magnitude, so a few high-norm tokens can crowd out the rest; this evens that out. Directions are preserved - only lengths change.
front_scale	FLOAT	1.000–2	Scales the first third of the ACTIVE region. 1.0 = off.
mid_scale	FLOAT	1.000–2	Scales the middle third of the ACTIVE region. 1.0 = off.
end_scale	FLOAT	1.000–2	Scales the last third of the ACTIVE region. 1.0 = off. Note these thirds are token positions, not meanings - the model assigns no fixed role to prompt position, so treat this as a coarse emphasis dial, not semantic targeting.
report	BOOLEAN	true	Print detected layout, token counts and before/after norms to the console. On by default because a conditioning transform you cannot inspect is indistinguishable from a placebo.

Outputs (1)

Name	Type	Description
CONDITIONING	CONDITIONING	—