LoadEmbedding

LoadEmbedding saves you from remembering ComfyUI's embedding syntax by hand. Instead of typing (embedding:filename:1.0) into your prompt yourself and hoping you spelled the filename right, this node gives you a dropdown of whatever embeddings are actually installed, a weight slider, and stitches the result onto whatever text you feed it.

Worth knowing up front, though: embeddings are a shrinking part of the ecosystem in 2026. They were once the go-to trick for compact negative prompts - EasyNegative and friends packing dozens of "avoid this" keywords into one small file - but the community has mostly moved to LoRAs for anything that needs real expressive power, and the reason is architectural, not just fashion: an embedding is a vector trained against one specific text encoder's vocabulary, and it simply doesn't bind to anything on the newer LLM-encoded models (Qwen3, Mistral) that a lot of 2026-era checkpoints use. It still works fine on SD 1.5, SDXL, and SDXL-lineage checkpoints like Illustrious or Pony - that's where essentially all new embeddings are still being published - but don't expect one to do anything on a Flux 2, Z-Image, or similar model.

How it works

You pick a file from embedding, set its weight, and optionally feed in existing prompt text via text. The node appends the embedding reference at your chosen weight onto that text and returns the combined string, ready to encode.

The inputs and outputs that matter

• text - optional-in-spirit but required by the schema; the prompt text you want the embedding appended to. Leave it empty if you just want the embedding reference on its own.
• embedding - a dropdown populated from whatever's in your embeddings folder. If it's empty when you open the node, that's not a bug - it means you haven't installed any embeddings yet (see below).
• weight - -2 to 2, step 0.1, default 1. Standard embedding weighting; higher pushes the effect harder, negative inverts it, though in practice most people stay close to 1.

One output: text, the finished string - feed it straight into a CLIPTextEncode, Textbox, or any of this pack's other text nodes.

How to install it

• ComfyUI Manager: search "ComfyUI-Chibi-Nodes", install, restart.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/chibiace/ComfyUI-Chibi-Nodes/, restart.

The node itself needs nothing extra, but it's only useful once you actually have embeddings on disk - download a .safetensors or .pt textual inversion (from CivitAI or elsewhere), drop it into ComfyUI/models/embeddings/, and restart or refresh ComfyUI so the dropdown picks it up.

The pack overall is a small personal utility collection - the README calls these "experimental nodes" and admits version 2 shipped "with more bugs" - worth keeping in mind as general calibration, though this node's logic (build a string, append a reference) is low-risk regardless.

Common issues

If embedding shows an empty dropdown, you don't have any embedding files in your models folder yet - that's the fix, not a bug in the node.

The bigger trap, and the one worth genuinely understanding rather than just working around: if you're running an embedding on a model whose text encoder it wasn't trained against, ComfyUI won't warn you - it just silently ignores what it can't apply. Your generation runs, nothing errors, and the embedding simply has zero effect. If you suspect this is happening, the reliable test is to rename the embedding file to something meaningless (or swap the weight to 0 vs 1 and compare on a fixed seed) - if the output doesn't change either way, the embedding isn't actually being read, and whatever effect you thought you saw was the filename acting as ordinary prompt text rather than the trained vector doing anything.