Nodes/LoRA Power-Merger ComfyUI/PM LoRA Merger (Mergekit)
ComfyUI Node

PM LoRA Merger (Mergekit)

Core LoRA merger using Mergekit algorithms. Merges decomposed LoRA components using the selected merge method (TIES, DARE, SLERP, etc.). Processes layers in parallel using ThreadPoolExecutor for performance. Inputs: - method: Merge algorithm configuration from method nodes - components: Decomposed LoRA tensors from LoRA Decompose node - strengths: Per-LoRA weight multipliers (strength_model for UNet, strength_clip for CLIP) - lambda_: Global scaling factor applied to final merged result (0-1) - merge_clip: Whether to include CLIP layers in the merge (default: True) Outputs: - LoRABundle: Merged LoRA ready for application or saving - MergeContext: Reusable merge configuration for batch operations

By larsupb·Created 2 years ago·Updated 9 days ago· 75
PM LoRA Merger (Mergekit)
  • method
  • components
  • strengths
  • lora
  • merge_context
lambda_1.00
spectral_norm_scale0.00
merge_cliptrue
device
dtype
CategoryLoRA PowerMerge

Inputs (8)

NameTypeDefaultDescription
methodMergeMethod
componentsLoRATensorsThe decomposed components of the LoRAs to be merged.
strengthsLoRAWeightsThe weights of the LoRAs to be merged.
lambda_FLOAT1.000–1Lambda value for scaling the merged model.
spectral_norm_scaleFLOAT0.000–100Spectral norm regularization. 0=disabled. >0=scale max Lipschitz constant to this value. Prevents any single layer from dominating due to large weight magnitudes. Common values: 0.5-2.0 (conservative), 1.0 (neutral), 2.0-5.0 (stronger effects).
merge_clipBOOLEANtrueWhether to merge CLIP (text encoder) layers. If False, CLIP layers are excluded from the output and LoRA Apply will not modify the CLIP model.
deviceCOMBO2 options: cuda, cpu
dtypeCOMBO3 options: float16, bfloat16, float32

Outputs (2)

NameTypeDescription
loraLoRABundle
merge_contextMergeContext