UCLA

Image-to-image translation is a fascinating and rapidly evolving field in computer vision and artificial intelligence. The problem involves transformation from an input image into an output image, while preserving certain semantic or structural information. This technology enables machines to convert data from one domain into another, offering a wide range of applications like artistic rendering and image restoration. In recent years, this field has garnered significant attention and research efforts, motivated by its potential to revolutionize various industries like entertainment and healthcare. In this dissertation, we address image-to-image translation challenges through a dual lens: cross-domain translation and cross-dimension translation. To be more precise, we present efficient and scalable approaches capable of accomplishing multi-domain translation within a unified framework. Additionally, we introduce an innovative 3D reconstruction method capable of generating three-dimensional representations from single 2D images. Through comprehensive experimentation on diverse datasets spanning multiple modalities, our findings not only validate the efficiency and effectiveness of our proposed methods but also signify a promising technological solution for facilitating efficient cross-domain and cross-dimension translation tasks.