Reproduce “Birth and Death of a Rose” 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
Chen Geng, Yunzhi Zhang, Shangzhe Wu, Jiajun Wu
This paper from Stanford University introduces a two-stage framework to generate temporally evolving 3D objects with dynamic geometry, reflectance, and texture, referred to as "temporal object intrinsics." The approach distills 4D-consistent signals from pre-trained 2D foundation models, producing high-fidelity, physically-based assets that enable novel view synthesis and relighting.
Key insights
- Distilling 4D-consistent temporal object intrinsics from 2D foundation models is achievable by creating a "Neural Template" that serves as a 4D control signal for generative processes.
- A hybrid 4D representation, combining K-Planes for low-frequency temporal consistency and Neural Graphical Primitives (NGP) for high-frequency details, effectively captures dynamic object properties.
- Bridging the domain gap between coarse 4D geometry renderings and natural image distributions, using a consistency model, is crucial for generating effective control signals for 2D diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.