Reproduce “Leveraging Pretrained Diffusion Models for Zero-Shot Part Assembly” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruiyuan Zhang, Qi Wang, Jiaxiang Liu, Yu Zhang, Yuchi Huo, Chao Wu
Researchers from Zhejiang University and North China Electric Power University developed a zero-shot method for 3D part assembly by leveraging pre-trained diffusion models as discriminative guides for an iterative optimization process. This approach assembles complex 3D objects without task-specific training, often outperforming existing supervised and unsupervised methods.
Key insights
- Pre-trained diffusion models can be repurposed as discriminative guides for assembly by treating them as density estimators that indicate object plausibility.
- The inherent prior knowledge within diffusion models about object structures and part relationships can be leveraged without task-specific training.
- Assembly can be framed as an optimization problem where parts are moved to maximize the log-density of the assembled point cloud, guided by the diffusion model's score function.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.