CWB Merge Text Encoders (3 Models)

The three-input version of the CWB text encoder merger. It takes three text encoder files from models/text_encoders and produces one consensus-blended encoder, where the blend adapts per tensor instead of using a single fixed ratio. Three inputs is where the median consensus earns its keep - a lone outlier encoder can't drag the center around the way it can with a mean.

How it works

Model A is the anchor: output metadata, tensor name/shape anchoring, and the preservation reference for anything missing or incompatible. Models B and C are equal-prior contributors. For each tensor, CWB computes an element-wise mean or median consensus, measures each of the three encoders' cosine similarity to that consensus, rejects below-threshold contributors, powers and normalizes the surviving weights, and produces the weighted sum (with optional norm rescaling). Text-encoder tensors always use fixed coordinate alignment, so no alignment toggle - the layers line up by position, which is correct for a fixed-architecture encoder.

Output keys are the union of all inputs; tensors missing from the anchor but present elsewhere are merged from the sources available; mismatch_mode decides between skip, zeros, and error for incompatible tensors. With three encoders, try a median preset (robust_medn) before the default mean - one out-of-family encoder on a layer is exactly the failure mode median handles best.

Inputs that matter

• execution_mode - MERGE writes the file; DOCUMENTATION ONLY returns the CWB reference and opens nothing.
• model_a / model_b / model_c - your three encoders. A anchors metadata.
• cwb_preset (default balanced_mean) - dense presets; robust_medn recommended for three inputs.
• output_filename (default cwb_merged_3_text_encoder) - written to models/text_encoders.
• save_dtype / override_dtype - dtype for generated tensors.
• 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.
• 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 detail). Load the result with a CLIP/Text Encoder Loader and use it for prompt conditioning as normal.

Why you'd reach for it

Three-way encoder merges are rare but real - combining two style-specific CLIPs with a base encoder, for example, to get prompt understanding that none of the three alone provides. It's a batch node: run it, then test the resulting file in a normal workflow. If you want to preview the similarity math first, run CheckpointModelAnalysis or the text-encoder analysis on the pairs.

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 - and expect the occasional experimental edge case.