Reproduce “BloomScene: Lightweight Structured 3D Gaussian Splatting for Crossmodal Scene 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

Xiaolu Hou, Mingcheng Li, Dingkang Yang, Jiawei Chen, Ziyun Qian, Xiao Zhao, Yue Jiang, Jinjie Wei and 2 more

BloomScene introduces a lightweight 3D Gaussian Splatting framework for generating structured 3D scenes from text or images. The approach yields superior visual fidelity and geometric consistency while reducing model storage overhead by up to 5.8 times compared to existing 3DGS methods.

Key insights

  • Integrating pixel-level, distribution-level, and smoothness constraints on depth maps effectively regularizes 3D Gaussian Splatting, improving geometric consistency.
  • Exploiting structural context around grouped 3D Gaussians (anchors) through hash feature grids and dynamic quantization allows for substantial model compression without sacrificing quality.
  • A progressive generation approach that carefully aligns newly generated content with existing scene structures helps maintain coherence and reduce cumulative errors across multiple views.

Summary by alphaXiv.

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