# DiffuScene: Scene Graph Denoising Diffusion Probabilistic Model for Generative Indoor Scene Synthesis

**Authors:** Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, Matthias Nießner  
**arXiv:** [2303.14207](https://arxiv.org/abs/2303.14207)  
**Date:** 2023-03-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.14207v1/

## Key points

- Traditional methods in indoor 3D scene synthesis rely on optimization techniques with predefined constraints based on room design rules, object category distributions, and human-object interactions.
- Recent advancements in deep generative models show promise in learning scene priors from large-scale datasets.
- GAN-based methods are used for implicit fitting of scene distributions through adversarial training, while VAE-based methods explicitly approximate scene distributions for better generative diversity.
- Auto-regressive models have been explored to predict objects sequentially but may struggle to capture relative attributes between objects and accumulate prediction errors over time.
- The introduced novel scene graph denoising diffusion probabilistic model generates 3D instance properties stored in a fully-connected scene graph and retrieves similar object geometries based on various attributes such as location, size, orientation, semantic information, and geometry features.
- By utilizing a diffusion model to determine placements and types of 3D instances, the model enables diverse applications including scene completion, arrangement, and text-conditioned synthesis.
- Experimental results on the 3D-FRONT dataset demonstrate that the model outperforms state-of-the-art methods by synthesizing more physically plausible and diverse indoor scenes.
- Extensive ablation studies further validate the effectiveness of the design choices made in the development of the scene diffusion model.

## Summary

In the realm of indoor 3D scene synthesis, traditional methods have often relied on optimization techniques with predefined constraints based on room design rules, object category distributions, and human-object interactions. However, these approaches are time-consuming and limited in their ability to capture complex scene arrangements. Recent advancements in deep generative models have shown promise in learning scene priors from large-scale datasets. Some approaches utilize GAN-based methods for implicit fitting of scene distributions through adversarial training, while VAE-based methods explicitly approximate scene distributions for better generative diversity. Auto-regressive models have also been explored to predict objects sequentially; however, they may struggle to effectively capture relative attributes between objects and can accumulate prediction errors over time. In this context,  introduces a novel scene graph denoising diffusion probabilistic model for generative indoor scene synthesis. This model generates 3D instance properties stored in a fully-connected scene graph and retrieves similar object geometries based on various attributes such as location, size, orientation, semantic information, and geometry features. By utilizing a diffusion model to determine placements and types of 3D instances, enables diverse applications including scene completion, arrangement, and text-conditioned synthesis. Experimental results on the 3D-FRONT dataset demonstrate that  outperforms state-of-the-art methods by synthesizing more physically plausible and diverse indoor scenes. Extensive ablation studies further validate the effectiveness of the design choices made in the development of the scene diffusion model. Overall, represents a significant advancement in indoor 3D scene synthesis by leveraging innovative probabilistic modeling techniques to generate realistic and varied scenes with greater efficiency and accuracy.