Core Concepts
Stein Score Distillation (SSD) incorporates flexible control variates derived from Stein's identity to effectively reduce the variance in the gradient estimation of score distillation, leading to improved text-to-3D generation quality and faster convergence.
Abstract
The paper presents SteinDreamer, a text-to-3D generation framework that addresses the high variance issue in score distillation techniques.
The key insights are:
The authors reveal that the variance of gradient estimation plays a crucial role in the performance of score distillation methods like SDS and VSD. They show that VSD exhibits lower variance compared to SDS, leading to better results.
Motivated by this observation, the authors propose Stein Score Distillation (SSD), which incorporates control variates derived from Stein's identity. This allows for flexible construction of control variates that can be highly correlated with the lifted image score, leading to significant variance reduction.
Specifically, the authors implement the control variate using a pre-trained monocular depth or normal estimator, which provides geometric guidance to the 3D optimization.
Extensive experiments demonstrate that SteinDreamer, the overall pipeline integrating SSD, consistently outperforms existing methods in both scene-level and object-level text-to-3D generation, producing sharper textures, more detailed geometries, and faster convergence.
Stats
The authors utilize 12 text prompts for scene-level generation and 20 text prompts for object-level generation.
Each scene or object is generated 3 times by each algorithm for evaluation.