[综述] Generative AI meets 3D: A Survey on Text-to-3D in AIGC Era

[综述] Generative AI meets 3D: A Survey on Text-to-3D in AIGC Era_第1张图片

论文|

改文章是23年5月27日挂在arxiv上,本文重点关注4.1节Text Guided 3D Avatar Generation、4.4节Text Guided 3D Shape Transformation和第5章Discussion


Text Guided 3D Avatar Generation

DreamAvatar

DreamAvatar: Text-and-Shape Guided 3D Human Avatar Generation via Diffusion Models

https://arxiv.org/abs/2304.00916生成姿态可控的高质量3D人体avatar,包含以下几个部分:

  • Trainable NeRF: 预测3D点的密度和颜色特征;
  • Pre-trained text-to-image diffusion model:提供2D自监督信息;
  • SMPL model:提供粗位姿和形态指导和生成;
  • dual-space design:包含canonical space和observation space,这两个空间与NeRF相关,将canonical space的纹理和几何信息传递到目标位姿的avatar;

该文章基于显示模型(SMPL)建模,缺点是缺少纹理细节,优点是可控性强。是目前的主流思路,还包括:

  • HeadSculpt: Crafting 3D Head Avatars with Text

DreamFace

DreamFace: Progressive Generation of Animatable 3D Faces under Text Guidance

文本引导的定制化3D人像生成,能够产生逼真的3D人像,包含以下几个部分:

  • Score Distillation Sampling (SDS): optimize subtle translations and normals
  • dual-path mechanism: 产生中性外观;
  • two-stage optimization: 增强用于细粒度合成的紧凑先验,并改善定制化能力

AvatarCraft: Transforming Text into Neural Human Avatars with Parameterized Shape and Pose Controls​​​​​​​

MotionCLIP

AvatarCLIP

Text Guided 3D Shape Transformation

传统方法对3D模型的编辑较为繁琐。因此,文本引导的3D模型编辑很有前景。

Instruct-NeRF2NeRF

Instruct 3D-to-3D

SKED

TextDeformer

你可能感兴趣的:(人工智能,AIGC)