Our contributions can be summarized as:• We propose to optimize NeRF as 3D self-prior, whereadaptive viewpoint sampling and boundary integrity metric are suggested to alleviate inconsistent generation.• We propose a control-based score distillation to maintain geometries in self-prior, where conditional LoRA andweighted score are presented to stabilize the optimization.• Our two-stage framework, namely DreamControl, cangenerate high-quality 3D content in text-to-3D generation, and the cont
(Optional:) Please add any files, screenshots, or other information here.
(Required) What is this issue most closely related to? Select one.
Description
Our contributions can be summarized as:• We propose to optimize NeRF as 3D self-prior, whereadaptive viewpoint sampling and boundary integrity metric are suggested to alleviate inconsistent generation.• We propose a control-based score distillation to maintain geometries in self-prior, where conditional LoRA andweighted score are presented to stabilize the optimization.• Our two-stage framework, namely DreamControl, cangenerate high-quality 3D content in text-to-3D generation, and the cont
(Optional:) Please add any files, screenshots, or other information here.
(Required) What is this issue most closely related to? Select one.
Choose One
Internal issue ID
5ba69adf-791e-46b7-a70e-62fa8dbaf5c4
Paper URL
https://arxiv.org/html/2312.06439v2
Browser
Chrome/123.0.0.0
Device Type
Desktop