Reproduce “CasaGPT: Cuboid Arrangement and Scene Assembly for Interior Design” 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

Weitao Feng, Hang Zhou, Jing Liao, Li Cheng, Wenbo Zhou

CASAGPT introduces a new framework for 3D indoor scene synthesis, employing a cuboid-based object representation and an autoregressive Transformer model with rejection sampling to minimize object collisions. The method consistently produces more physically plausible and visually coherent scenes, achieving up to a 31% higher non-intersection rate compared to prior methods.

Key insights

  • A fine-grained, cuboid-based object representation, rather than bounding boxes, is crucial for effectively modeling spatial relationships and preventing mesh-level object collisions.
  • Combining an autoregressive generative model with a post-hoc iterative rejection sampling mechanism effectively distills optimal layout policies for collision-free scene generation.
  • Cleaning noisy training datasets, which often contain pre-existing object intersections, is essential for improving the physical plausibility and quality of synthesized scenes.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.