CWB Merge Embeddings (3 Models)

The three-input version of the CWB embedding merger - the answer to "can I combine three textual-inversion vectors into one file?" Yes, and with per-token consensus weighting rather than a blind average. It's the most niche merger in the pack, but for anyone who collects multiple embeddings of the same concept it's genuinely handy.

How it works

Model A is the anchor (metadata, preservation reference). Models B and C are equal-prior contributors. For each token vector, CWB computes a consensus - with three contributors the median becomes a robust center that a single outlier embedding can't drag - measures each embedding's similarity to it, and weights contributions by that similarity. Embeddings use the longest compatible first dimension as their alignment reference, and the default preset (balanced_idx_mean) aligns by index; if you believe the token order differs between files, the similarity-alignment presets (balanced_sim_mean, robust_sim_medn) pair tokens by cosine instead.

Output is written to models/embeddings and used like any embedding - reference embedding:filename in a prompt or wire it through an embedding-aware text encoder. mismatch_mode handles missing or incompatible tokens: skip preserves the anchor, zeros inserts zero contributions, error aborts.

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 embeddings. A anchors metadata.
• cwb_preset (default balanced_idx_mean) - the embedding registry; _medn variants use median consensus, _sim variants use similarity alignment.
• output_filename (default cwb_merged_3_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, force_clear_cache, process_device - the usual UEL streaming, flushing, and device controls.

Outputs

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

Why you'd reach for it

Consolidating three partial embeddings of the same concept - each trained on a different slice of examples - into one vector is the realistic case. The output is small and the merge is fast, so it's low-risk to try: merge, prompt with embedding:, and judge by eye. If you want to preview token similarity first, EmbeddingModelAnalysis from the same pack has the alignment tooling.

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.