VETIM: Expanding the Vocabulary of Text-to-Image Models only with Text

Text-to-image models, such as Stable Diffusion, can generate high-quality images from simple textual prompts. With methods such as Textual Inversion, it is possible to expand the vocabulary of these models with additional concepts, by learning the vocabulary embedding of new tokens. These methods have two limitations: slowness of optimisation and dependence on sample images. Slowness mainly stems from the use of the original text-to-image training loss, without considering potential auxiliary supervision terms. Relying on sample images enables learning new visual features but restricts the vocabulary expansion to concepts with pre-existing images. In response, we introduce a novel approach, named VETIM, which takes only a textual description of the concept as input. It expands the vocabulary through supervision only at the text encoder output, without accessing the image-generation part, making it faster at optimisation time. It also does not copy visual features from sample images. Our method can be used directly for applications that require a concept as a single token but do not require learning new visual features. Our approach shows that a mere textual description suffices to obtain a single token referring to a specific concept. To show the effectiveness of our method, we evaluate its performance subjectively and through objective measures. The results show that our approach is effective in expanding the vocabulary of text-to-image models without requiring images.