This paper explores facial attribute transformation, able to modify features like age, gender, and hairstyle in digital images or videos. Facial attribute transformation has multiple applications like entertainment, cosmetics, and security. Traditional methods typically train separate models for each facial attribute independently which means to train as many transformers as facial attributes, requiring extensive computational resources. In this work, we introduce LMAT, a unified, efficient and compact architecture that can learn multiple facial attributes simultaneously using a single Latent Transformer model. During inference, LMAT performs multi-attribute transformations in just one step that makes it fast and efficient. Our method generates more realistic results and better preserves individual identities, all while being eight times faster in training and inference than the current state-of-the-art methods. Furthermore, subjective analysis with 33 users indicates our approach delivers superior performance in facial attribute transformation. The project source code is available and will be made public upon the publication of this manuscript.