Save Token Embeddings

Let's get the big misconception out of the way first: Save Token Embeddings is not a trainer. You aren't teaching the model a new concept, and you aren't doing textual inversion. The node takes a prompt you already have, runs it through the CLIP text encoder, and writes the resulting token vectors to a small .safetensors file in your embeddings folder. Think of it as freezing a prompt into a reusable file - a precomputed conditioning snippet you can summon with embedding:filename in any future prompt.

That's genuinely useful, even if it sounds underwhelming. Bake your house-style prompt once and reuse it everywhere. Collapse a long negative prompt into one token (the EasyNegative trick, without the 40-word preamble). Share a prompt as a 50 KB file instead of a 200-word text wall. Because the file is just vectors, it's deterministic - same model, same file, same conditioning, no drift from retyping.

How it works

The node tokenizes your text with the current CLIP's tokenizer, then walks every text-encoder component the loaded model actually has - clip_l / clip_g on SDXL, t5xxl on Flux, the Qwen or Mistral parts on newer LLM-encoded checkpoints. For each component it runs the token IDs through the encoder's process_tokens, keeps only the real (non-padding) tokens, and concatenates them into one tensor. The result is saved as a multi-key safetensors file straight into ComfyUI/models/embeddings, where ComfyUI's built-in embedding loader can find it next time you write embedding:name in a text encoder.

Note that this is the unweighted version - any (word:1.3) style weights in your text are tokenized but not applied to the saved vectors. If you want weights baked in (which matters on Flux-class models that discard them), use the pack's Save Weighted Embeddings node instead.

The inputs that matter

Only four, and you'll touch two of them most of the time:

• clip - wire in any CLIP loader. The node figures out which sub-encoders exist.
• text - the prompt to freeze. Multiline, supports dynamic prompts.
• slice_bos_eos (default off) - leave it off and the file includes the BOS/EOS markers, so using it in a prompt reproduces the original conditioning almost exactly. Turn it on and it strips BOS/EOS from clip_l/clip_g and EOS from the T5 parts, giving you a leaner vector sequence that slots into a prompt more like a plain token. Start with off; slice later if the embedding fights your prompt.
• filename_prefix (default token_embeds) - files save as token_embeds_00001.safetensors, auto-incremented so you never overwrite.

Where it saves, and how you use it

Files land in ComfyUI/models/embeddings/, not your output folder. The numbering means the file's name is token_embeds_00001 - reference it in any text encoder as embedding:token_embeds_00001. There's no output port to wire; the node's job is done the moment it writes the file.

Installing the pack

This one is refreshingly painless. The pack (silveroxides/ComfyUI_EmbeddingToolkit) has zero dependencies beyond ComfyUI itself - no requirements.txt, no model downloads, nothing to pip install. Grab it via ComfyUI Manager (search "ComfyUI_EmbeddingToolkit") or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI_EmbeddingToolkit

Restart ComfyUI and you're done. It's a small personal utility from silveroxides - better known as the maintainer of the well-regarded Chroma-GGUF quants - so don't expect a big community; expect something that just works.

Gotchas

The big one is the encoder lock. An embedding is bound to the text encoder it was created with: an SDXL CLIP file won't mean anything on a Qwen3 or Mistral model, and a Flux/T5 file won't work on SDXL. The UI won't warn you - it silently ignores what it can't apply, which is how people convince themselves incompatible embeddings are working. Save files only for the model family you actually generate with.

Also worth knowing: the README carries an explicit disclaimer - creating embeddings with this and uploading them to CivitAI under Early Access is called out as deceptive. Free uploads are fine, but don't be that person.