CWB Merge LoRAs (3 Models)

Three LoRAs into one file, blended by consensus per layer

By silveroxides·Created about a year ago·Updated about 24 hours ago· 14

cwb_config

output_filename
documentation
cwb_report

◄execution_mode▾►

◄model_a▾►

◄model_b▾►

◄model_c▾►

◄cwb_presetbroad_sim_medn_rn_softcb►

◄mismatch_modeskip►

◄output_filenamecwb_merged_3_lora►

◄save_dtype▾►

◄process_device▾►

◄exclude_patterns►

◄discard_patterns►

◄glob_patternsfalse►

◄lazy_loadtrue►

◄force_clear_cachetrue►

◄override_dtypefalse►

◄include_1d_diffsfalse►

The three-input version of the CWB LoRA merger - the one that lets you collapse a three-LoRA stack into a single file. If you've been running two or three LoRA loaders at fractional strengths and fighting the overcooked look, this is the node that does the combination in advance, per layer, without a single global strength guess.

How it works

Same Consensus-Weighted Blending engine as CWBLoRATwoMerger, with a third equal-prior contributor. Model A anchors metadata, naming, and preservation; Models B and C contribute. LoRA down/up pairs are treated as one logical layer each, and for every layer CWB computes a consensus (mean or median), measures each of the three LoRAs' cosine similarity to it, and weights contributions by similarity. Three inputs is where the median consensus really shines - if one LoRA is an outlier on some layer, median keeps it from dragging the blend. The LoRA presets default to similarity-based alignment (LoRA rows have no fixed semantic order), and the default broad_sim_medn_rn_softcb is a solid middle ground; strongfocus_sim_medn_rn_dsc_softcb pushes the differences harder.

Inputs that matter

execution_mode - MERGE writes the file; DOCUMENTATION ONLY returns the CWB reference without loading.
model_a / model_b / model_c - your three LoRAs. A anchors metadata.
cwb_preset (default broad_sim_medn_rn_softcb) - the LoRA registry; consider a median variant, which is the default here.
include_1d_diffs (default off) - merges the 1D tensors (alphas, biases) in fp32 instead of preserving Model A's. Leave off unless you need it.
output_filename (default cwb_merged_3_lora) - written to models/loras.
save_dtype / override_dtype - dtype for generated factors.
mismatch_mode - skip / zeros / error for missing or incompatible groups.
exclude_patterns / discard_patterns / glob_patterns - regex (or glob) lists.
lazy_load (default on), force_clear_cache (default on), process_device - the usual streaming, flushing, and CUDA-with-CPU-fallback.

Outputs

output_filename (the written file), documentation (the CWB reference), and cwb_report (per-layer detail). Load the merged file with a normal LoRA Loader.

The realistic use case

Three LoRAs each doing one job - a character, a style, a quality/detail boost - that you want as a single coherent file rather than a three-loader stack. Merging to one file also tends to reduce the brittle parameter juggling that stacked LoRAs force on you. Run LoRAModelAnalysis on the pairs first if you want to know how much they overlap.

Install

From Model Utility Toolkit (silveroxides/ComfyUI-ModelUtils). ComfyUI Manager → search "Model Utility Toolkit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils

Restart ComfyUI. Real dependency: unifiedefficientloader (UEL). Keep ComfyUI current - newer extension API, experimental CWB nodes.

CategoryModelUtils/Merging

Inputs (17)

Name	Type	Default	Description
execution_mode	COMBO		MERGE writes a new safetensors file. DOCUMENTATION ONLY returns the CWB reference without loading or merging inputs.
model_a	COMBO		Primary contributor and preservation anchor. Supplies output metadata and anchors shared tensor names and shapes.
model_b	COMBO		Second equal-prior contributor. CWB derives its effective per-vector influence from consensus similarity.
model_c	COMBO		Third equal-prior contributor. CWB derives its effective per-vector influence from consensus similarity.
cwb_preset	COMBO	broad_sim_medn_rn_softcb	Use-case preset. Name suffixes expose alignment, consensus, norm rescaling, DSC, soft comfort bandpass, and prefix preservation. A connected CWB Config overrides it completely.
mismatch_mode	COMBO	skip	For missing or incompatible anchored inputs: skip preserves the anchor, zeros inserts a zero contribution where possible, and error aborts. A lone secondary-only tensor is copied unchanged.
output_filename	STRING	cwb_merged_3_lora	Filename without extension. The result is atomically written to this model category under ComfyUI's models directory.
save_dtype	COMBO		Requested dtype for generated floating tensors. Participating FP32 inputs keep a result FP32 unless Override Dtype is enabled.
process_device	COMBO		Device used for per-layer FP32 CWB arithmetic. A CUDA out-of-memory error retries only the affected layer on CPU.
exclude_patterns	STRING		One pattern per line. Matching layers are preserved from the anchor instead of merged. Uses regex unless Glob Patterns is enabled.
discard_patterns	STRING		One pattern per line. Matching tensors or logical LoRA groups are omitted from the output. Uses regex unless Glob Patterns is enabled.
glob_patterns	BOOLEAN	false	Interpret exclude and discard entries as shell-style glob patterns instead of regular expressions.
lazy_load	BOOLEAN	true	Use UEL low-memory loading so tensors are read and released per work unit instead of retaining the complete inputs in RAM.
force_clear_cache	BOOLEAN	true	Run Python garbage collection and clear the CUDA allocator cache before each layer. Reduces retained memory but can substantially slow merging.
override_dtype	BOOLEAN	false	Force generated tensors to save_dtype; guarded tensors and enabled 1D direct diffs are exempt.
include_1d_diffs	BOOLEAN	false	CWB-merge 1D direct diffs as FP32. Disabled preserves Model A.
cwb_configopt	CWB_CONFIG		Optional settings from CWB Custom Configuration. When connected, it completely overrides the selected preset.

Outputs (3)

Name	Type	Description
output_filename	*	—
documentation	STRING	—
cwb_report	STRING	—