An embedding is a small file that adds a new custom keyword to a checkpoint's vocabulary — trained through a process called textual inversion — without changing any of the model's actual weights.
Textual inversion training finds a new point in CLIP's text-embedding space that represents a specific style or subject shown in a handful of training images, then saves just that point as a tiny file. Typing the embedding's trigger word in your prompt calls up that point directly, without retraining or modifying the checkpoint itself.
Where You'll See It
Stored in ComfyUI/models/embeddings, and used by typing its trigger word straight into a CLIP Text Encode (Prompt) node — no dedicated loader node is required the way LoRA needs one.
Quick Example
EasyNegative is a widely used negative embedding for SD1.5 — typing it into your negative prompt calls up a bundle of "bad quality" concepts in one word, instead of typing a long list of quality-flaw terms every time.
If a downloaded embedding doesn't seem to change anything, check that it was trained for the same base model family (SD1.5, SDXL, etc.) as your checkpoint — embeddings rarely transfer cleanly across architectures.
Common Mix-Up
Embedding vs. LoRA: an embedding only adds a new keyword vector — it never touches the checkpoint's weights, which limits how much it can change. A LoRA actually modifies the model's weights (in a small, efficient way), which is why LoRAs can produce a much stronger, more reliable effect than an equivalent embedding.
Tip: Looking to train your own custom concept instead? See our ComfyUI LoRA guide for the more modern approach.
Frequently Asked Questions
Type its trigger word directly into a CLIP Text Encode node — usually the filename without the extension, sometimes written as embedding:name depending on the custom nodes you have installed. ComfyUI Manager’s autocomplete extension can help surface installed embeddings as you type.
Less than they used to be — LoRA has largely replaced embeddings for adding new subjects or styles, since it can change more of the model’s behavior. Embeddings are still popular specifically for negative-prompt shortcuts, like a single keyword that bundles a long list of quality-flaw terms.
Not reliably. An embedding is trained against one specific model’s CLIP text-encoder space, so it usually works best on the checkpoint it was trained for, or close relatives of the same base model family.
See It In Action
Want a stronger custom effect than an embedding?
Our LoRA guide covers loading and stacking LoRAs for stronger, more reliable custom styles.
