This presentation is divided into three parts. Firstly, we will go through a series of advancements that have redefined the landscape of image and video generation, such as GenTron (CVPR’24), Video DiT (ICLR’24), and MotionCtrl (SIGGRAPH’24) and the PixArt series—PixArt-alpha (ICLR’24), PixArt-delta (arXiv:2401.05252), and PixArt-sigma (arXiv:2403.04692). Secondly, we introduce how to unify perception and generation capacity in a single multimodal LLM, such as the instance-level Vision Language Models for image understanding and generation tasks (RegionGPT, CVPR’24). These models are distinct from traditional multimodal LLMs fine-tuned with image-text pairs, which often face challenges in achieving detailed instance-level visual concepts. Thirdly, building on the success of large multimodal models in high-level understanding, we design a multimodal code generation framework, RoboCodeX (ICML’24), crafted to convert task plans into precise robotic actions, ensuring adaptability across diverse scenarios. Our approach seeks to seamlessly integrate high-level cognitive processing with practical robotic applications, paving the way for enhanced robotic autonomy and versatility.