GALA3D: Towards Text-to-3D Complex Scene Generation via Layout-guided Generative Gaussian Splatting

Zhou et al. introduce GALA3D, a compositional TextTo3D pipeline that targets multi-object scenes rather than the single-object outputs typical of DreamFusion-era generators. The motivating observation is that most SDS-based pipelines collapse when asked to generate scenes with several distinct entities, because a single monolithic 3D field cannot easily enforce separation, scale, and interaction constraints between objects referred to in the prompt.

The pipeline begins by querying a large language model to translate a free-form scene description into an explicit layout: a list of named objects with approximate bounding boxes and qualitative relationships. This LLM-generated layout serves as the prior for a layout-guided GaussianSplatting representation, where Gaussians are organised into per-object groups and the optimisation is constrained by per-object geometric priors derived from the bounding boxes. Each object can be optimised semi-independently while still influencing its neighbours through the global rendering loss.

An instance-scene compositional optimisation mechanism, driven by a conditioned diffusion prior, then refines the scene jointly. This stage not only sharpens textures and shapes but also adjusts the initial coarse layout that came from the LLM, so that scale, contact, and relative positions become physically and semantically plausible. The result is a compositional pipeline that supports both high-quality scene-level generation and per-object controllable editing: the user can move, resize, or swap objects in the layout and re-optimise. GALA3D is a contemporary of LucidDreamer in the open-domain scene generation literature; the two papers differ mainly in whether scene structure comes from an inpainting-driven point cloud or from an LLM-emitted layout.