ComfyUI Node

LC Krea2 LoRA Editor

Loads a Krea 2 LoRA with per-block, attention / MLP and spike edits, applies it to the model and outputs the edited LoRA for LC LoRA Save. The heatmap shows each layer's strength before (outline) and after (fill) the edits.

By lonecatone23·Created 11 days ago·Updated 2 days ago· 1
LC Krea2 LoRA Editor
  • model
  • model
  • edited_lora
◄lora_name▾►
◄strength1.00►
◄presetFull►
◄blocks1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1►
◄attention1.00►
◄mlp1.00►
◄outside_blocks1.00►
◄spike_softenoff►
◄spike_limit2.0►
◄blocks_in—►
CategoryLC ModelBuilder/Krea2

Inputs (11)

NameTypeDefaultDescription
modelMODELKrea 2 model. The edited LoRA is applied to it, like a LoRA loader.
lora_nameCOMBOLoRA to edit, from models/loras.
strengthFLOAT1.00-4–4Overall LoRA strength, same as a LoRA loader. Baked into the saved LoRA.
presetCOMBOFullFull = the LoRA as it is. Blocks X-Y down = that group at 0.3: face lock sits in different blocks for each LoRA, so try each group on a fixed seed batch and keep the one that frees the face (Lip Bite: 16-21, Amateur V4: 4-9). Tame = blocks 4-15 at 0.7, which with spike_soften directions stopped Lip Bite changing the whole image. Dragging the graph switches to Custom.
blocksSTRING1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 128 block multipliers, 0 to 2 (1 = unchanged). Drag the bars on the node.
attentionFLOAT1.000–2Multiplier for the attention layers (q, k, v, o, gate) in every block.
mlpFLOAT1.000–2Multiplier for the MLP layers (gate, up, down) in every block. Most of a Krea 2 LoRA's weight usually sits here.
outside_blocksFLOAT1.000–2Multiplier for layers outside the 28 blocks (text fusion, first, last), if the LoRA has any. The heatmap note says how many.
spike_softenCOMBOofflayers = pull any layer that is far stronger than the same layer in other blocks down to the limit. directions = inside each layer, flatten the one or two directions that dominate it (plain lora_A / lora_B files only). both = both.
spike_limitFLOAT2.01–10How strong a spike may be before it is softened. layers: times the median of that layer across the 28 blocks. directions: times the layer's second strongest direction (well-behaved LoRAs sit around 2). Lower = more softening.
blocks_inoptSTRINGA 28-block curve from another node, such as LC Krea2 LoRA Block Scan's blocks output. When wired it replaces the preset and the bars.

Outputs (2)

NameTypeDescription
modelMODELThe model with the edited LoRA applied.
edited_loraLC_LORAThe edited LoRA, for LC LoRA Save.