Reproduce “LucidDreaming: Controllable Object-Centric 3D Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaoning Wang, Ming Li, Chen Chen
LucidDreaming introduces a plug-and-play pipeline for controllable object-centric 3D generation, leveraging LLMs and novel rendering techniques to enable precise spatial and numerical control over multi-object scenes. The system significantly improves object placement and quantity accuracy across various text-to-3D frameworks and uniquely allows for new object insertion into existing NeRF scenes.
Key insights
- Large Language Models can effectively translate complex spatial and numerical constraints from natural language into concrete 3D bounding box layouts for generative models.
- Localized rendering and density initialization are critical for achieving distinct, controllable, and accurately placed objects in multi-object 3D generation without fusion or omission.
- A modular, plug-and-play design allows for seamless integration with existing SDS-based 3D generation frameworks, enhancing their control capabilities without extensive retraining.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.