# HierOctFusion: Multi-scale Octree-based 3D Shape Generation via Part-Whole-Hierarchy Message Passing

**Authors:** Xinjie Gao, Bi'an Du, Wei Hu  
**arXiv:** [2508.11106](https://arxiv.org/abs/2508.11106)  
**Date:** 2025-08-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.11106v1/

## Key points

- Complexity of 3D data poses a challenge in content generation
- Existing methods overlook semantic part hierarchies of 3D objects
- HierOctFusion is a part-aware multi-scale octree diffusion model introduced to enhance hierarchical feature interaction
- HierOctFusion incorporates a cross-attention conditioning mechanism for more accurate shape generation
- Authors construct a specialized 3D dataset with part category annotations for training and evaluation
- HierOctFusion outperforms previous techniques in shape quality and computational efficiency

## Summary

, , , , 

In the realm of 3D content generation, the complexity of 3D data poses a significant challenge. While <kw>octree-based diffusion models</kw> have shown promise in balancing efficiency and quality through hierarchical generation, they often overlook crucial aspects. Specifically, existing methods tend to treat 3D objects as holistic entities, neglecting their semantic part hierarchies and thereby limiting generalization. Moreover, the computational expense associated with holistic high-resolution modeling contrasts with the inherent sparsity and hierarchy of real-world objects, suggesting a need for layered generation approaches. In response to these observations, Xinjie Gao, Bi'an Du, and Wei Hu introduce HierOctFusion—a novel <kw>part-aware multi-scale octree diffusion model</kw> designed to enhance hierarchical feature interaction for generating fine-grained and sparse object structures. This innovative approach incorporates a cross-attention conditioning mechanism that integrates part-level information into the generation process. By enabling effective propagation of semantic features across hierarchical levels—from individual parts to complete objects—HierOctFusion facilitates more accurate and detailed shape generation. To support their research efforts, the authors construct a specialized 3D dataset with part category annotations obtained using a pre-trained segmentation model. This dataset serves as a valuable resource for training and evaluating HierOctFusion's performance in comparison to existing methods. Through a series of experiments, the authors demonstrate that HierOctFusion outperforms previous techniques in terms of both shape quality and computational efficiency. Overall, HierOctFusion represents a significant advancement in the field of 3D shape generation by addressing key limitations in current approaches and leveraging <kw>part-aware multi-scale modeling</kw> to achieve superior results.