Nodes/ComfyUI_LC_ModelBuilder_nodes/LC Krea2 Semantic Vector Tuner
ComfyUI Node

LC Krea2 Semantic Vector Tuner

Amplify or attenuate specific regions of a single Krea2 checkpoint's weights. Not a merge. Reuses Arthemy SDXL Suite's own tuning mechanism (bake existing patches, scale, re-patch the delta via ModelPatcher.add_patches) pointed at Krea2's real block/region prefixes instead of SDXL's UNet layout. 3 zone sliders cover the 28 uniform blocks (with an optional per-block override string for finer control); the 10 named, architecturally distinct regions each get their own slider.

By lonecatone23·Created about 18 hours ago·Updated about 18 hours ago· 0
LC Krea2 Semantic Vector Tuner
  • model
  • model
  • info
◄modeSoft Value►
◄base_scale1.00►
◄low_blocks1.00►
◄mid_blocks1.00►
◄high_blocks1.00►
◄blocks_override►
◄first1.00►
◄tmlp1.00►
◄txtmlp1.00►
◄tproj1.00►
◄txtfusion_layer01.00►
◄txtfusion_layer11.00►
◄txtfusion_projector1.00►
◄txtfusion_refiner01.00►
◄txtfusion_refiner11.00►
◄last1.00►
CategoryLC ModelBuilder/Krea2

Inputs (17)

NameTypeDefaultDescription
modelMODELSingle checkpoint to amplify/attenuate by region. This is not a merge — no second model.
modeCOMBOSoft ValueSoft Value dampens the sliders so 0-2 stays safe (values above 1.0 land around 1.0-1.13x). Real Value applies each slider as a literal multiplier.
base_scaleFLOAT1.000–2Fallback multiplier for any weight that isn't one of the 28 blocks or the 10 named regions.
low_blocksFLOAT1.000–2Scales blocks.0. - blocks.8. (low zone).
mid_blocksFLOAT1.000–2Scales blocks.9. - blocks.17. (mid zone).
high_blocksFLOAT1.000–2Scales blocks.18. - blocks.27. (high zone).
blocks_overrideSTRINGOptional comma list of exactly 28 multipliers, one per block. Overrides low_blocks/mid_blocks/high_blocks for the 28 blocks when present. Named regions below are unaffected — they always come from their own sliders.
firstFLOAT1.000–2first. — earliest input/embedding weights, before block 0.
tmlpFLOAT1.000–2tmlp. — timestep-embedding MLP (timestep conditioning).
txtmlpFLOAT1.000–2txtmlp. — text-embedding MLP (text conditioning).
tprojFLOAT1.000–2tproj. — timestep projection layer.
txtfusion_layer0FLOAT1.000–2txtfusion.layerwise_blocks.0. — first text/image fusion block.
txtfusion_layer1FLOAT1.000–2txtfusion.layerwise_blocks.1. — second text/image fusion block.
txtfusion_projectorFLOAT1.000–2txtfusion.projector. — projects fused text features into the main stream.
txtfusion_refiner0FLOAT1.000–2txtfusion.refiner_blocks.0. — first fusion refiner block.
txtfusion_refiner1FLOAT1.000–2txtfusion.refiner_blocks.1. — second fusion refiner block.
lastFLOAT1.000–2last. — final output layer, after block 27.

Outputs (2)

NameTypeDescription
modelMODEL—
infoSTRING—