Reproduce “GASP: Gaussian Avatars with Synthetic Priors” 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
Jack Saunders, Charlie Hewitt, Yanan Jian, Marek Kowalski, Tadas Baltrusaitis, Yiye Chen, Darren Cosker, Virginia Estellers and 3 more
GASP introduces a method for creating photorealistic, real-time, animatable 360-degree digital human avatars from limited monocular input data. It leverages a generative prior trained on a large-scale synthetic dataset and a three-stage fitting process to achieve high-quality free-viewpoint rendering.
Key insights
- Large-scale synthetic datasets with precise annotations can effectively train robust 3D generative priors for digital humans, overcoming real-world data limitations.
- Utilizing learnable, semantic per-Gaussian feature vectors enables efficient scaling of the prior model and facilitates consistent appearance updates across observed and unseen regions during fitting.
- A multi-stage fitting process with prior regularization effectively bridges the synthetic-to-real domain gap, allowing a prior trained on synthetic data to generalize to real users while maintaining fidelity in unseen regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.