ComfyUI-WordEmbeddings

Utilities for exploring and composing word embeddings in ComfyUI using gensim models. Use these nodes to inspect neighborhoods, traverse between concepts, and solve free‑form word “equations”.

Included nodes (core):

• WordEmbeddingsLoader — pick a gensim model by name.
• WordEmbeddingsExplorer — list nearest neighbors to a word.
• WordEmbeddingsInterpolator — sample words along the path between two words.
• WordEmbeddingsEquation — evaluate expressions like king - man + woman.

Default model: glove-wiki-gigaword-50 (small & fast).

---

Installation

If you’re on ComfyUI Windows Portable, use its embedded Python.

1. Place this repo at:

ComfyUI/custom_nodes/ComfyUI-WordEmbeddings

2. Install dependencies:

Windows (PowerShell)

cd .\ComfyUI_windows_portable\python_embeded
.\python.exe -m pip install --upgrade pip
.\python.exe -m pip install -r ..\ComfyUI\custom_nodes\ComfyUI-WordEmbeddings
equirements.txt

macOS / Linux

pip install -r ComfyUI/custom_nodes/ComfyUI-WordEmbeddings/requirements.txt

3. Start ComfyUI. Nodes appear under “ComfyUI-WordEmbeddings”.

Model downloads: On first use, gensim downloads the selected model to your user cache (e.g., ~/gensim-data or %USERPROFILE%\gensim-data). To override, set GENSIM_DATA_DIR before launching ComfyUI.

---

Quick start

1. Drop WordEmbeddingsLoader → choose glove-wiki-gigaword-50 (fast).
2. Feed its model_name into:
   • Explorer (type a word, get top‑k neighbors), or
   • Interpolator (set word_a, word_b, choose stops), or
   • Equation (enter an expression like paris - france + italy).

Save any STRING output using your preferred text saver node.

---

Nodes

1) WordEmbeddingsLoader

Select a gensim model via dropdown.

Inputs

• model_name (choice) — list of installed/available gensim models (default: glove-wiki-gigaword-50 if present).

Outputs

• model_name (STRING) — pass to other nodes.

---

2) WordEmbeddingsExplorer

Return nearest neighbors for a single query word.

Inputs

• model_name (STRING) — from Loader.
• word (STRING) — e.g., king.
• k (INT) — neighbors to return (default 10).

Outputs

• words_csv (STRING) — comma‑separated neighbors.
• debug_csv (STRING) — word:similarity pairs.

Example

• word = king, k = 10
  → words_csv: prince,queen,monarch,...
  → debug_csv: prince:0.78,queen:0.77,...

---

3) WordEmbeddingsInterpolator

Sample along the path from word_a to word_b and report the nearest word at each interior step.

Inputs

• model_name (STRING) — from Loader.
• word_a, word_b (STRING) — endpoints (e.g., king, queen).
• stops (INT) — interior samples (no endpoints; default 12).
• (optional) topn (INT) — candidate pool size for nearest lookup (default 1000).
• (optional) method (choice) — slerp (default) or lerp.

Outputs

• words_csv (STRING) — samples only, comma‑separated.
• debug_txt (STRING) — per‑step diagnostics (cos to A/B/mid, etc.).

Notes

• slerp (spherical linear interpolation) generally respects direction on the embedding hypersphere better than lerp.

---

4) WordEmbeddingsEquation

Solve free‑form vector arithmetic from a text expression.

Expression syntax

• +, -, plus, minus
• Numeric coefficients: 2*woman, 2.5 man
• Ignores punctuation like ?, ->
• Anything after = or equals is ignored
  (e.g., king - man + woman = ? works the same as without = ?)

Inputs

• model_name (STRING) — from Loader (default glove-wiki-gigaword-50).
• expression (STRING) — e.g., king - man + woman.
• k (INT) — number of results to return (default 10).
• (optional) topn_pool (INT) — pool size from similar_by_vector (default 2000).
• (optional) normalize_terms (BOOL) — L2‑normalize each term vector before summing (default True; recommended).
• (optional) allow_inputs (BOOL) — include input words in results (default False).

Outputs

• words_csv (STRING) — top‑k candidates.
• debug_txt (STRING) — parse details, used/missing terms, cosine scores.

Examples

• king - man + woman → often queen
• paris - france + italy → often rome
• son - man + woman → often daughter

---

Tips & troubleshooting

• Speed: glove-wiki-gigaword-50 is the quickest to download and load. Larger models (e.g., GoogleNews 300) can be richer but heavier.
• Case: Nodes attempt a lowercase fallback if the exact case is OOV.
• First-run slowness: The first time a model is selected, gensim will download it; subsequent runs use the cache.
• Top‑N pools: If an interpolated sample returns odd neighbors, try raising topn / topn_pool.

---

License

Released under the MIT License. See the included LICENSE file.