핵심 개념
DreamControl proposes a two-stage framework to optimize 3D generation, focusing on geometry consistency and texture fidelity.
초록
DreamControl introduces a novel approach to text-to-3D generation by optimizing coarse NeRF scenes as 3D self-prior and then refining objects with control-based score distillation. The method addresses issues like viewpoint bias and overfitting, resulting in high-quality 3D content. It can be applied to user-guided generation and 3D animation tasks.
The paper discusses the challenges in current 2D-lifting techniques for 3D generation, highlighting the Janus problem caused by inconsistent geometry. By proposing adaptive viewpoint sampling and boundary integrity metrics, DreamControl ensures consistent generated priors. The control-based optimization guidance enhances downstream tasks like user-guided generation and 3D animation.
Furthermore, DreamControl's two-stage framework optimizes NeRF as a self-prior to maintain geometry consistency while generating detailed textures. Extensive experiments demonstrate its superiority in geometry consistency, texture fidelity, and text-to-3D generation tasks.
통계
"DreamControl can generate high-quality 3D content in terms of both geometry consistency and texture fidelity."
"DreamControl surpasses competing methods in all evaluation metrics."