The view inconsistency problem in score-distilling text-to-3D generation, also known as the Janus problem, arises from the intrinsic bias of 2D diffusion models, which leads to the unrealistic generation of 3D objects. In this work, we explore score-distilling text-to-3D generation and identify the main causes of the Janus problem. Based on these findings, we propose two approaches to debias the score-distillation frameworks for robust text-to-3D generation. Our first approach, called score debiasing, involves gradually increasing the truncation value for the score estimated by 2D diffusion models throughout the optimization process. Our second approach, called prompt debiasing, identifies conflicting words between user prompts and view prompts utilizing a language model and adjusts the discrepancy between view prompts and object-space camera poses. Our experimental results show that our methods improve realism by significantly reducing artifacts and achieve a good trade-off between faithfulness to the 2D diffusion models and 3D consistency with little overhead.

翻译：得分蒸馏式文本到3D生成中的视角不一致问题（亦称Janus问题）源于二维扩散模型的内在偏差，导致3D物体的生成缺乏真实性。本研究深入探究得分蒸馏式文本到3D生成机制，揭示了Janus问题的主要成因。基于这些发现，我们提出两种针对得分蒸馏框架的去偏方法以实现稳健的文本到3D生成。第一种方法称为评分去偏，通过在优化过程中逐步增大二维扩散模型估计得分的截断值。第二种方法称为提示去偏，利用语言模型识别用户提示与视角提示之间的冲突词，并调整视角提示与物体空间相机位姿之间的差异。实验结果表明，我们的方法通过显著减少伪影提升了生成真实感，并在忠实于二维扩散模型与保持3D一致性之间实现了良好权衡，且计算开销极低。