Disentangled 3D Scene Generation with Layout Learning
This addresses the challenge of creating editable 3D content for applications like virtual reality or gaming, though it appears incremental as it builds on existing NeRF and text-to-image methods.
The paper tackles the problem of generating 3D scenes that are disentangled into component objects without supervision, using a pretrained text-to-image model to ensure valid configurations, resulting in successful decomposition enabling new text-to-3D capabilities.
We introduce a method to generate 3D scenes that are disentangled into their component objects. This disentanglement is unsupervised, relying only on the knowledge of a large pretrained text-to-image model. Our key insight is that objects can be discovered by finding parts of a 3D scene that, when rearranged spatially, still produce valid configurations of the same scene. Concretely, our method jointly optimizes multiple NeRFs from scratch - each representing its own object - along with a set of layouts that composite these objects into scenes. We then encourage these composited scenes to be in-distribution according to the image generator. We show that despite its simplicity, our approach successfully generates 3D scenes decomposed into individual objects, enabling new capabilities in text-to-3D content creation. For results and an interactive demo, see our project page at https://dave.ml/layoutlearning/