Sustainable artificial intelligence (AI) has become an urgent issue with the advancements in deep learning in recent years. From the data perspective, the recent large-scale foundation models are facing the bottleneck of data in network training. From the computation perspective, the increasing model parameters demand increasing GPU resources and power consumption. In this talk, I will share our recent research that attempts to address the AI sustainability challenge through image generation and cross-domain transfer. Specifically, image generation learns to compose, translate, and edit images, aiming for automated generation of high-fidelity and self-annotated images that can be effectively applied for training deep neural networks. Cross-domain transfer investigates unsupervised domain adaptation, unsupervised model adaptation, and domain generalization, aiming to better leverage previously learnt knowledge or collected data for mitigating the high demand for GPU while handling various downstream applications and tasks.