Hyper Remap Krea2 Edit (CCN)

Four-phase remapping pipeline fused with the krea2_edit image-grounded encode. comma (,) = text replacement. arrow (->) = token embedding blend. fat arrow (=>) = concept direction nudge (cosine mode when grounded). double tilde (~~) = additive residual delta (pooled when grounded). Per-pair (b:X, s:X, t:X, sx:X, tx:X) overrides in parentheses take precedence over globals. Connect image (and optionally image_b) to ground the encode; empty remappings = plain grounded encode.

By valkymaera·Created 3 months ago·Updated 4 days ago· 1

clip
image
image_b

conditioning
original_prompt
modified_prompt

◄text—►

◄remappings# string replace: find, replace # token remap: source -> target # token remap: source -> target (0.8) # concept remap: source => target # concept remap: source => target (b:0.8, s:2.0, t:0.1) # delta remap: base ~~ subtracted # delta remap: base ~~ subtracted (b:0.5, s:1.0, sx:2.0) ►

◄blend1.000►

◄sharpness1.00►

◄threshold0.000►

◄grounding_px768►

◄normalize_deltatrue►

◄case_sensitivetrue►

◄debugfalse►

CategoryCCN/conditioning

Inputs (12)

Name	Type	Default	Description
clip	CLIP		—
text	STRING		—
remappings	STRING	# string replace: find, replace # token remap: source -> target # token remap: source -> target (0.8) # concept remap: source => target # concept remap: source => target (b:0.8, s:2.0, t:0.1) # delta remap: base ~~ subtracted # delta remap: base ~~ subtracted (b:0.5, s:1.0, sx:2.0)	—
blend	FLOAT	1.000-100–100	Default blend for all operators. For ->: lerp between original and remapped embeddings. For => and ~~: magnitude of the nudge vector. >1 overshoots, negative inverts direction. Per-pair (b:X) overrides take precedence.
sharpness	FLOAT	1.00-100–100	Default incoming-conditioning sharpness for => and ~~. Controls how sharply positions are weighted by their cosine similarity to the source/base concept. 0 = uniform across all positions. Higher = concentrated on most-similar positions. Negative = favour least-similar positions. Ignored by ->. Per-pair (s:X) overrides take precedence.
threshold	FLOAT	0.000-1–1	Default incoming-conditioning threshold for => and ~~. Masks out positions whose similarity weight falls below this value after sharpness is applied. 0 = all positions eligible. Ignored by ->. Per-pair (t:X) overrides take precedence.
imageopt	IMAGE		Source image to ground the encode on (krea2_edit semantic path). Leave disconnected for a text-only encode.
image_bopt	IMAGE		2nd reference (subject) for multi-ref LoRAs; vision blocks in training order: scene first, subject second.
grounding_pxopt	INT	7680–4096	Cap the longest side fed to Qwen3-VL; 0 = native resolution.
normalize_deltaopt	BOOLEAN	true	L2-normalise the delta tensor before blending for ~~ entries. When on, blend has a consistent magnitude regardless of how different the two prompts are. When off, larger semantic differences produce stronger effects at the same blend value.
case_sensitiveopt	BOOLEAN	true	Case sensitivity for string replacement pairs.
debugopt	BOOLEAN	false	Print phase diagnostics to console.

Outputs (3)

Name	Type	Description
conditioning	CONDITIONING	—
original_prompt	STRING	—
modified_prompt	STRING	—