Reproduce “Disco4D: Disentangled 4D Human Generation and Animation from a Single Image” 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
Hui En Pang, Shuai Liu, Zhongang Cai, Lei Yang, Tianwei Zhang, Ziwei Liu
Disco4D introduces a Gaussian Splatting framework that generates disentangled, animatable 4D human models from a single image, separating the SMPL-X body from clothing and dynamic elements. This approach achieves high-fidelity 3D reconstruction and animation, enabling precise editing and outperforming existing methods in quantitative metrics and user preference.
Key insights
- Disentangling the human body from clothing using distinct Gaussian representations (fixed SMPL-X for body, flexible for clothes) enables independent manipulation and animation.
- Embedding clothing Gaussians onto the SMPL-X mesh with learnable identity encodings allows for fine-grained categorization, animation, and localized editing of individual assets.
- Combining robust human priors (SMPL-X) with generative models (diffusion priors for texture, learned deformations for dynamics) significantly enhances fidelity and realism, particularly for occluded areas and dynamic movements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.