Faster, Smaller, and Simpler Model for Multiple Facial Attributes Transformation

Abstract

There are many existing models that are capable of changing hair color or changing facial expressions. These models are typically implemented as deep neural networks that require a large number of computations in order to perform the transformations. This is why it is challenging to deploy on a mobile platform. The usual setup requires an internet connection, where the processing can be done on a server. However, this limits the application’s accessibility and diminishes the user experience for consumers with low internet bandwidth. In this paper, we develop a model that can simultaneously transform multiple facial attributes with lower memory footprint and fewer number of computations, making it easier to be processed on a mobile phone. Moreover, our encoder–decoder design allows us to encode an image only once and transform multiple times, making it faster as compared to the previous methods where the whole image has to be processed repeatedly for every attribute transformation. We show in our experiments that our results are comparable to the state-of-the-art models but with $4\times $ fewer parameters and $3\times $ faster execution time.