Reproduce “Zero-P-to-3: Zero-Shot Partial-View Images to 3D Object” 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
Yuxuan Lin, Ruihang Chu, Zhenyu Chen, Xiao Tang, Lei Ke, Haoling Li, Yingji Zhong, Zhihao Li and 4 more
Zero-P-to-3 is a training-free framework that reconstructs complete 3D objects from partial-view images by combining 3D Gaussian Splatting with multi-prior diffusion models. The method achieves high fidelity in observed regions and accurate reconstruction of unseen areas, resulting in PSNR 18.74, SSIM 0.8634, and LPIPS 0.0913 on the Objaverse dataset.
Key insights
- Combining 3D Gaussian Splatting with multi-prior diffusion effectively leverages rich local information while synthesizing missing regions for 3D reconstruction.
- Multi-prior score fusion (geometric rendering, image restoration, multi-view diffusion) overcomes conditional independence assumptions in multi-view diffusion, ensuring consistency across visible and invisible regions.
- An iterative refinement strategy with rotated view sampling progressively expands coverage and refines the 3DGS model using generated novel views as supervision.
- Time-dependent weight scheduling in score fusion allows structural guidance to dominate early and texture refinement to become prominent in later diffusion stages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.