CWB Merge Models (3 Models)
Three UNets/DiTs, one consensus-merged denoiser
- cwb_config
- output_filename
- documentation
- cwb_report
The three-input CWB merger for standalone diffusion models - UNet or DiT files in models/diffusion_models. It's the three-model sibling of CWBModelTwoMerger, and with three contributors the consensus math gets meaningfully stronger, because the median becomes a genuinely robust center that a single outlier can't skew.
How it works
Model A anchors the merge: it supplies output metadata, anchors tensor names and shapes, and is the preservation reference for anything missing or incompatible. Models B and C are equal-prior contributors. For each tensor, the engine computes an element-wise consensus (mean or median per preset), measures each of the three models' cosine similarity to that consensus, and weights each contribution by its similarity - no global blend ratio anywhere. Output keys are the union of all inputs; tensors only present in secondary inputs are merged from the sources available or copied; mismatch_mode decides between skip (preserve anchor), zeros (zero contribution), and error (abort).
The practical tip for three inputs: try a median preset first. With three models, one of them being a lone outlier on some layer is common, and robust_medn (or diverse_medn_rn_dsc_softcb if you want the diversity bandpass) handles that far better than a mean that gets dragged toward the outlier.
The result is a single denoiser written to models/diffusion_models - load it with your architecture's UNet/DiffusionModel loader, then bring your own text encoder and VAE.
Inputs that matter
- execution_mode -
MERGEwrites the file;DOCUMENTATION ONLYreturns the CWB reference and opens nothing. - model_a / model_b / model_c - your three files. A anchors metadata and naming.
- cwb_preset (default
balanced_mean) - the dense-model presets; considerrobust_mednhere. - output_filename (default
cwb_merged_3_model) - written tomodels/diffusion_models. - save_dtype / override_dtype - dtype for generated tensors; fp32 inputs stay fp32 unless forced.
- exclude_patterns / discard_patterns / glob_patterns - regex (or glob) lists; excludes preserve matching layers from the anchor, discards drop them.
- lazy_load (default on) - UEL streaming; keep it on so three large files don't pile up in RAM.
- force_clear_cache (default on) - per-layer flushing; memory-safe, slower.
- process_device - CUDA with per-layer CPU retry on OOM.
Outputs
output_filename (the written file), documentation (the CWB reference), and cwb_report (per-layer merge detail worth reading after the first run).
Why you'd reach for it
A base, a style model, and a character/quality specialist - the classic three-way stack that per-layer consensus weighting handles better than any fixed ratio. It's a batch node: it writes a file and stops. Run DiffusionModelAnalysis on the pairs first if you want a preview of the similarity math.
Install
From Model Utility Toolkit (silveroxides/ComfyUI-ModelUtils), one install for the whole pack. 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 updated - newer extension API - and expect the author-flagged experimental edge cases on oddball files.
Inputs (16)
| 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 | balanced_mean | 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_model | 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. |
| 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 | — |