# DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior

**Authors:** Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, Yebin Liu  
**arXiv:** [2310.16818](https://arxiv.org/abs/2310.16818)  
**Date:** 2023-10-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.16818v2/

## Key points

- DreamCraft3D is a hierarchical 3D content generation method
- It produces high-fidelity and coherent 3D objects
- It addresses the consistency issue faced by existing methods
- It uses a 2D reference image to guide geometry sculpting and texture boosting
- DreamCraft3D employs score distillation sampling through a view-dependent diffusion model for coherently rendered geometries
- The authors propose Bootstrapped Score Distillation to boost texture fidelity
- They train a personalized diffusion model called Dreambooth on augmented renderings of the scene
- The optimization process involves alternating optimization of the diffusion prior and 3D scene representation
- The optimized 3D scene aids in training the scene-specific diffusion model for view-consistent guidance in 3D optimization.

## Summary

DreamCraft3D is a novel hierarchical 3D content generation method that produces high-fidelity and coherent 3D objects. It addresses the consistency issue faced by existing methods by leveraging a 2D reference image to guide the stages of geometry sculpting and texture boosting. To ensure coherently rendered geometries, DreamCraft3D employs score distillation sampling through a view-dependent diffusion model. This 3D prior combined with various training strategies prioritizes geometry consistency but compromises texture fidelity. To overcome this limitation, the authors propose Bootstrapped Score Distillation which specifically focuses on boosting texture. The authors train a personalized diffusion model called Dreambooth on augmented renderings of the scene to imbue it with 3D knowledge of the optimized scene and obtain view-consistent guidance for generating the final output. The optimization process involves an alternating optimization of the diffusion prior and 3D scene representation resulting in mutually reinforcing improvements. The optimized 3D scene aids in training the scene-specific diffusion model which offers increasingly view-consistent guidance for 3D optimization.