Textual inversion

Textual inversion is a lightweight way to teach a model something new. Instead of changing the checkpoint, it learns a brand-new "word" - a small file called an embedding - that points at a concept you trained from a handful of example images. You then drop that keyword into any prompt to summon the concept.

How it works

During training, the model's weights are frozen. The only thing that changes is a new text embedding - the numeric vector a keyword maps to - which is nudged until it reliably produces your target concept. Because nothing in the base model moves, the resulting file is tiny (often just a few kilobytes), far smaller than a LoRA or a full checkpoint.

Embeddings vs LoRA vs DreamBooth

• Textual inversion: learns a new keyword only. Tiny file, good for a consistent style or a simple object; limited capacity.
• LoRA: trains a small set of extra weights. More capacity for characters and detailed styles; larger file.
• DreamBooth: fine-tunes the base model itself. Most faithful, but produces a heavy file.

Why it matters

Embeddings are the cheapest, most shareable way to add a recurring concept to your prompts. A common use is a negative embedding - a keyword bundling many quality flaws so you can clean up output with a single token in your negative prompt.