Partverse
收藏资源简介:
Partverse是一个大规模的3D部件数据集,包含91,000个部件,用于训练名为CoPart的3D生成框架。该数据集由Objaverse通过自动网格分割和后续人工后注释收集而成。CoPart框架利用多个上下文部件潜在表示来表示3D对象,并能够同时生成一致的3D部件。此外,为了确保部件潜在表示的一致性,并利用基础模型强大的先验知识,论文提出了一种新的相互引导策略,以微调预训练的扩散模型,从而实现联合部件潜在去噪。CoPart在Partverse数据集上的训练实现了具有高可控性的基于部件的3D生成,并支持各种应用,包括部件编辑、关节对象生成和小场景生成。
Partverse is a large-scale 3D part dataset containing 91,000 parts, designed for training the CoPart 3D generative framework. This dataset is collected from Objaverse via automated mesh segmentation followed by manual post-annotation. The CoPart framework leverages multiple contextual part latent representations to represent 3D objects, and is capable of generating consistent 3D parts simultaneously. Furthermore, to ensure the consistency of part latent representations and exploit the strong prior knowledge of foundation models, this work proposes a novel mutual guidance strategy for fine-tuning pre-trained diffusion models, enabling joint part latent denoising. Training CoPart on the Partverse dataset achieves part-based 3D generation with high controllability, supporting various applications including part editing, articulated object generation, and small-scale scene generation.

- 1From One to More: Contextual Part Latents for 3D Generation香港科技大学, 香港中文大学, 商汤科技研究院, 上海人工智能实验室 · 2025年



