Compositional Generative

Compositional generative modeling focuses on building complex generative models by combining smaller, more specialized models, rather than relying on single, monolithic architectures. Current research emphasizes using diffusion models and neural fields, often within a graph-based framework, to represent and manipulate individual components of a scene, enabling control over object arrangement and attributes. This approach offers improved data efficiency, generalization to unseen combinations of elements, and enhanced controllability in applications such as image composition, 3D scene generation, and robotic manipulation planning, ultimately advancing the capabilities of generative AI systems.