Reproduce “Wukong's 72 Transformations: High-fidelity Textured 3D Morphing via Flow Models” 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
Minghao Yin, Yukang Cao, Kai Han
The WUKONG framework enables high-fidelity textured 3D morphing, generating smooth geometric and texture transitions between 3D objects from image or text prompts. It leverages a flow-based 3D generator and optimal transport for latent space interpolation, achieving an FID score of 4.01 and allowing fine-grained control over texture preservation during transformations.
Key insights
- Leveraging the deterministic and continuous properties of flow-based generative models provides a robust latent space for smooth 3D object manipulation.
- Solving an optimal transport barycenter problem with recursive initialization effectively navigates the data manifold, ensuring semantically coherent and geometrically stable latent interpolations.
- Decoupling and selectively blending texture features based on semantic similarity allows for fine-grained control over appearance changes, crucial for identity-preserving morphing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.