DreamScene: 3D Gaussian-based End-to-end Text-to-3D Scene Generation
Haoran Li, Yuli Tian, Kun Lan, Yong Liao, Lin Wang, Pan Hui, Peng Yuan Zhou
2025-07-31
Summary
This paper talks about DreamScene, a new system that can create detailed 3D scenes automatically just by understanding text descriptions or dialogues, allowing people to easily generate and control 3D environments.
What's the problem?
The problem is that making 3D scenes usually requires a lot of manual work, skill, and time, and existing AI methods struggle to keep the 3D world consistent and editable based only on text input.
What's the solution?
DreamScene solves this by using a step-by-step approach where it plans the scene, uses graphs to decide where objects should go, samples formation patterns to arrange things naturally, and adjusts the camera view progressively, all based on a special 3D Gaussian model that keeps everything coherent and editable.
Why it matters?
This matters because it makes creating 3D scenes much easier and faster for designers, game developers, and artists, letting them turn their ideas into 3D worlds with simple text instructions.
Abstract
DreamScene is an end-to-end framework that generates high-quality, editable 3D scenes from text or dialogue, ensuring automation, 3D consistency, and fine-grained control through a combination of scene planning, graph-based placement, formation pattern sampling, and progressive camera sampling.