SceneBloom: Agentic Indoor Environment Generation with Realistic and Diverse Layouts for Embodied Learning

TL;DR

SceneBloom generates realistic, diverse, and physically valid indoor 3D environments by combining diffusion-based layout sampling with a geometric compiler guided by spatial relations. It produces layouts with no mesh collisions or out-of-bounds objects and better matches real-world room distributions than prior methods.

Why It Matters

Realistic and varied indoor environments are essential for training embodied agents that generalize to real-world settings. SceneBloom enables higher-quality training simulations, leading to better-performing navigation policies when deployed in real spaces.

Key Findings

• SceneBloom achieves the highest visual quality and physical stability across five room types compared to existing methods.
• It produces layouts that are statistically closer to real-world room distributions than all baselines.
• Agents trained in SceneBloom-generated environments outperform those trained on real scans in held-out real-world tests.