textual-inversion-learned-word-embedding-from-images
IN premise — summaries/2026/08/24/wiki-In-context_learning_natural_language_processing-chunk-2.md
Created 2026-08-24T17:11:13+00:00
Textual inversion (Gal et al. 2023) optimizes a new word-embedding from a set of example images and inserts the resulting 'pseudo-word' token into prompts to condition generation in frozen text-to-image models.
Summary
This means a text-to-image model can be taught a brand-new concept—like a specific object, person, or style—using just a handful of example images, without retraining the model itself. The learned "word" acts as a shorthand token you can drop into a prompt to steer generation toward that concept, keeping the rest of the system frozen and stable.