Pair2Scene: Learning Local Object Relations for Procedural 3D Scene Generation

Forum topic · 小凯 · 2026-04-15

Summary

Pair2Scene is a procedural generation framework for creating high-fidelity 3D indoor scenes, proposed by Xingjian Ran, Shujie Zhang, Weipeng Zhong, Li Luo, and Bo Dai (arXiv:2604.11808, cs.CV). The work addresses two persistent challenges in indoor scene synthesis: data scarcity and the difficulty of modeling intricate spatial relationships. The authors observe that existing methods either fail to scale beyond their training distribution to dense scenes, or depend on LLMs/VLMs that lack precise spatial reasoning. Pair2Scene's core insight is that object placement is governed mainly by local dependencies between object pairs rather than redundant global distributions. The framework learns two types of inter-object relations: support relations, which follow physical hierarchy, and functional relations, which reflect semantic links. It integrates these learned local rules with scene hierarchies and physics-based algorithms. Experiments show the method can generate complex environments that go beyond the training data distribution while remaining physically and semantically plausible.

Overview

Research area: cs.CV
Authors: Xingjian Ran, Shujie Zhang, Weipeng Zhong, Li Luo, Bo Dai
Published: 2026-04-13
arXiv: 2604.11808

Abstract (original)

Generating high-fidelity 3D indoor scenes remains a significant challenge due to data scarcity and the complexity of modeling intricate spatial relations. Current methods often struggle to scale beyond training distribution to dense scenes or rely on LLMs/VLMs that lack the ability for precise spatial reasoning. Building on top of the observation that object placement relies mainly on local dependencies instead of information-redundant global distributions, in this paper, we propose Pair2Scene, a novel procedural generation framework that integrates learned local rules with scene hierarchies and physics-based algorithms.

Key Points

Problem: High-fidelity 3D indoor scene generation is hindered by data scarcity and the complexity of modeling intricate spatial relations.
Limitations of existing methods: They struggle to scale beyond the training distribution to dense scenes, or rely on LLMs/VLMs that lack precise spatial reasoning.
Core insight: Object placement depends mainly on local dependencies between object pairs, not on information-redundant global distributions.
Learned relations: The framework captures two types of inter-object relations:
Support relations, following physical hierarchy.
Functional relations, reflecting semantic links between objects.
Approach: Pair2Scene integrates learned local rules with scene hierarchies and physics-based algorithms.
Results: The method generates complex environments that extend beyond the training data distribution while maintaining physical and semantic plausibility.

---

*Auto-collected on 2026-04-15.*