Nodes/EasyLoRAMerger/🎨 Easy LoRA Merger
ComfyUI Node

🎨 Easy LoRA Merger

A ComfyUI node in LoRA/Experimental with 29 inputs and 5 outputs.

By Terpentinas·Created 6 months ago·Updated 3 months ago· 11
🎨 Easy LoRA Merger
  • model
  • clip
  • lora_data_a
  • lora_data_b
  • lora_data_c
  • lora
  • model
  • clip
  • output_path
  • forensic_report
methodlinear
density1.00
lora_a
lora_b
lora_c
weight_a1.00
weight_b1.00
weight_c1.00
blend_modeauto
energy_preservationtrue
balancing_modesafe
magnitude_scalingnone
active_thresholdtrue
uniqueness0.70
threshold0.00
blend0.50
deviceauto
precisionauto
batch_size32
streamingtrue
save_triggerfalse
filenametriple_merged
save_folder/tmp/ComfyUI/models/loras
metadata_modemerge_basic
CategoryLoRA/Experimental

Inputs (29)

NameTypeDefaultDescription
modelMODEL
clipCLIP
methodCOMBOlinear15 options: linear, ties_strict, ties_gentle, dare_lite, dare_rescale, subtract, +9
densityFLOAT1.000.1–1
lora_aoptCOMBO1 options: None
lora_boptCOMBO1 options: None
lora_coptCOMBO1 options: None
lora_data_aoptLORA
lora_data_boptLORA
lora_data_coptLORA
weight_aoptFLOAT1.00-5–5Strength of first LoRA
weight_boptFLOAT1.00-5–5Strength of second LoRA
weight_coptFLOAT1.00-5–5Strength of third LoRA
blend_modeoptCOMBOautoauto: Smart choice based on trainer metadata (match → dense, mismatch → active) | dense: Traditional weighted sum
energy_preservationoptBOOLEANtruePreserve energy distribution during merge (recommended). Disable for raw weighted sum.
balancing_modeoptCOMBOsafeAuto-weight-balancing: safe (hard-match energy on shared layers only, rank-independent), creative (preserve ratio with reduced magnitude), intensity (peak-energy detection for cross-concept merges), impact (intensity + sparsity correction for sparse-vs-dense merges like Anima), disabled (no adjustment).
magnitude_scalingoptCOMBOnoneSignal magnitude scaling before merging – scales LoRA B and C to match LoRA A's energy using RMS or top‑X% percentile.
active_thresholdoptBOOLEANtrueEnable active region detection (threshold value from config.py). Disable for dense-style behavior.
uniquenessoptFLOAT0.700.1–1For feature_mix: higher = preserve more unique features
thresholdoptFLOAT0.000–1For subtract: minimum magnitude to subtract
blendoptFLOAT0.500–1For magnitude: 0=strict, 1=blended
deviceoptCOMBOauto3 options: auto, cuda, cpu
precisionoptCOMBOauto4 options: auto, float32, bfloat16, float16
batch_sizeoptINT321–256Number of keys to process per batch. DeviceManager.suggest_batch_size() can auto-tune based on VRAM.
streamingoptBOOLEANtrueStream tensors to save VRAM
save_triggeroptBOOLEANfalse
filenameoptSTRINGtriple_merged
save_folderoptSTRING/tmp/ComfyUI/models/loras
metadata_modeoptCOMBOmerge_basicHow to handle metadata from source LoRAs

Outputs (5)

NameTypeDescription
loraLORA
modelMODEL
clipCLIP
output_pathSTRING
forensic_reportSTRING