HetScene: Heterogeneous Diffusion for Dense Indoor Scene Generation
A new research paper introduces HetScene, a heterogeneous two-stage generation framework for indoor scene synthesis. The work addresses limitations in existing deep learning methods that treat all objects as homogeneous instances, which fails for dense arrangements with complex spatial dependencies. HetScene decomposes objects into primary and secondary categories based on their structural roles, then decouples layout generation into Structural Layout Generation (SLG) and Contextual Layout stages. The framework aims to improve controllability and physical plausibility for high-fidelity simulation environments in embodied AI. The paper is published on arXiv under identifier 2605.13586.
Key facts
- HetScene is a heterogeneous two-stage generation framework for indoor scene synthesis.
- It decomposes objects into primary and secondary objects based on structural roles.
- The framework decouples layout generation into Structural Layout Generation (SLG) and Contextual Layout stages.
- Existing deep learning methods treat all objects as homogeneous instances.
- Homogeneous methods struggle with dense object arrangements and complex spatial dependencies.
- HetScene aims to improve controllability and physical plausibility.
- The work targets high-fidelity simulation environments for embodied AI.
- The paper is available on arXiv with identifier 2605.13586.
Entities
Institutions
- arXiv