CWB Merge Embeddings (2 Models)

Embeddings - the small textual-inversion vector files - don't get merged much, and when they do, most people just average the files blindly. This is the CWB version, which means the blend adapts per token instead of being a flat mean. It's niche, but if you've ever wanted the concept understanding of two embeddings in one file, this is the node.

How it works

Model A is the anchor (metadata and preservation reference); Model B is the second equal-prior contributor. For each token vector, CWB computes a consensus, measures each embedding's similarity to it, and weights contributions by that similarity. Embeddings are a special case in the CWB family for a real reason: their token rows don't have a guaranteed semantic order, so the merge uses the longest compatible first dimension as its alignment reference, and the preset defaults to index alignment (balanced_idx_mean) with a similarity-alignment alternative (balanced_sim_mean) available if you believe the two files' tokens don't line up by position.

The output is an embedding file written to models/embeddings. You load it the way you load any embedding - reference embedding:your_filename in a prompt (or wire it through an embedding-aware text encoder). Missing or shape-incompatible tokens are handled by mismatch_mode: skip preserves the anchor, zeros inserts a zero contribution, error aborts.

Inputs that matter

• execution_mode - MERGE writes the file; DOCUMENTATION ONLY returns the CWB reference without loading.
• model_a / model_b - the two embeddings.
• cwb_preset (default balanced_idx_mean) - the embedding preset registry. balanced_sim_mean and robust_sim_medn switch to cosine-similarity alignment for files whose tokens may be in different orders; the _medn variants use median consensus.
• output_filename (default cwb_merged_2_embedding) - written to models/embeddings.
• save_dtype / override_dtype - dtype for generated tensors.
• exclude_patterns / discard_patterns / glob_patterns - regex (or glob) lists.
• lazy_load (default on), force_clear_cache (default on), process_device - the usual.

Outputs

output_filename (the written file), documentation (the CWB reference), and cwb_report (per-token merge detail).

Why you'd reach for it

Two embeddings trained on the same concept by different people, and you want the combined understanding. Or you've been using two related embeddings (say, a broad style and a narrow refinement) and want one file to reference. It's a batch node - merge, then test by prompting with embedding: and eyeballing.

Install

Part of 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.