Visual content creation includes a wide range of vision tasks, such as text-to-image generation, photo animation, novel view synthesis, digital avatar creation, etc. The past few years have witnessed the bloom of this area with diverse task settings and pipelines. Despite all these successful attempts, there seems to be many trade-offs, like the one between visual quality and controllability and the one between reconstruction fidelity and rendering efficiency. Such trade-offs also become the central issues when one opts for a data representation or a learning paradigm. It is noteworthy that elegant solutions usually begin with figuring out the best choice.
Visual Content Creation
沈宇军