4D Scene Generation

4D scene generation studies how to build dynamic 3D scenes that remain coherent as objects move and viewpoints change, from image, video, text, or action inputs. Common approaches lift video diffusion into dynamic Gaussian, mesh, or point-cloud representations; generate structured scene tokens autoregressively; or combine object-level motion with physics simulation. Language and trajectory controls are increasingly used to support editing, interaction, and embodied simulation. A central challenge is balancing visual quality with geometric, temporal, and physical consistency, especially for long sequences and unseen viewpoints where training data and reliable evaluations are limited.