What is a Textual Inversion embedding?

A Textual Inversion embedding is a small file that teaches Stable Diffusion a new concept — a style, character or quality fix — and binds it to a trigger word. Typing that word in your prompt invokes the learned concept without changing the base model.

Do I write the embedding name or use angle brackets?

In Automatic1111/Forge you just type the embedding's filename as a plain word, for example easynegative. In ComfyUI the standard syntax is embedding:name. Either way the name must match the file in your embeddings folder exactly.

Where do I put a negative embedding?

Negative embeddings such as easynegative or badhandv4 go in the negative prompt box. They steer the model away from the artifacts they were trained on, like deformed hands or watermarks.

How do I change an embedding's strength?

Wrap it in weight syntax just like any token — (easynegative:0.8) in WebUI, or use the same parenthesis weighting in ComfyUI. Lower the weight if the embedding is overpowering the rest of the prompt.

What is the Textual Inversion Embedding Guide?

Guide for using Textual Inversion embeddings in Stable Diffusion prompts. Covers the difference between bare-name and angle-bracket syntax across WebUI and ComfyUI, where to place positive and negative embeddings, and how to adjust their weight. Generates copy-ready prompt snippets. It runs free in your browser on Gera Tools, with nothing uploaded.

Textual Inversion Embedding Guide

Name: Textual Inversion Embedding Guide
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

Use embeddings without the syntax confusion

Textual Inversion embeddings are tiny files that teach Stable Diffusion a new style, character or quality fix and bind it to a trigger word. They are one of the cheapest ways to upgrade your output — easynegative alone cleans up a huge range of artifacts. The only thing that trips people up is the syntax, which differs between Automatic1111 and ComfyUI. This guide generates the exact token to paste, in the right prompt box, at the weight you want.

How embeddings work and where they go

An embedding does not modify the model’s weights. Instead it adds a new “word” to the text encoder’s vocabulary that maps to a learned concept. To use it you simply reference its filename (minus the extension):

Automatic1111 / Forge — type the bare name: easynegative, my-art-style. The UI auto-detects files in your embeddings/ folder.
ComfyUI — use the prefix form: embedding:easynegative. The file lives in models/embeddings/.

Placement matters:

Positive embeddings (styles, subjects, quality boosters) go in the positive prompt, usually near the front for style or with your subject.
Negative embeddings (easynegative, badhandv4, ng_deepnegative) go in the negative prompt — they describe what to avoid.

Understanding what Textual Inversion actually does

Textual Inversion (the training technique, not just the resulting files) works by finding a new token in the CLIP text encoder’s embedding space that best activates the target concept. During training, the model’s weights stay frozen and only the new token’s embedding vector is optimised. This is why embeddings are small (a few KB) and architecture-specific: the learned vector is tied to the specific CLIP encoder used by SD 1.5 or SDXL — it is literally coordinates in that encoder’s learned vocabulary space.

When you type the trigger word in a prompt, the text encoder replaces it with this learned vector before the diffusion process sees the text. Everything else in the prompt is processed normally.

Tips on weighting and stacking

Tune the weight. If an embedding dominates, dial it down: (my-style:0.7). If it is too weak, raise toward 1.2. Standard Automatic1111 weighting syntax works on embeddings the same way it works on regular tokens.
Don’t over-stack negatives. Three or four well-chosen negative embeddings cover most artifacts. Adding ten or fifteen often produces flat, desaturated images because the negative prompt begins overriding desirable features too.
Names are exact and case-sensitive on some systems. Copy the filename precisely — a typo means the token is silently treated as an unknown word rather than loading the embedding.
SD 1.5 vs SDXL. Embeddings trained on SD 1.5’s CLIP encoder will not load on an SDXL model (which uses a different, dual-encoder architecture) and vice versa. Verify the model architecture before downloading.

Common negative embeddings and what they target

Embedding	What it addresses
easynegative	General low-quality outputs, deformed anatomy, bad composition
badhandv4	Specifically malformed hands and fingers — a known SD weakness
ng_deepnegative	Deep negative conditioning for anatomy and quality issues
verybadimagenegative	Watermarks, text in images, compressed-looking outputs

These are widely used community embeddings. Results vary by base model and sampling settings, so test each one on your typical prompts and disable any that change the look more than you want.

Prompt placement in Automatic1111

For positive embeddings that define an overall style, placing them early in the prompt (before the subject description) gives them stronger influence. For quality booster embeddings like detailed, photo-realistic, placing them mid-prompt works well. Experiment with order to find what feels right for your specific model and style target.