Controlling embedding spaces with text-conditioned transformations

The paper proposes a method to control multimodal embedding spaces (e.g., CLIP) by applying text-conditioned transformations, enabling semantic similarity manipulation and zero-shot classification adjustments without retraining.