Textual Inversion

Textual Inversion is a technique that teaches a frozen text-to-image diffusion model a new concept, such as an object or style, from just three to five example images by learning a new pseudo-word in the model's embedding space. That learned word can then be composed into ordinary prompts to generate the concept in new scenes, a lightweight alternative to full fine-tuning. It is implemented in Diffusers and Stable Diffusion UIs.