Reproduce “GAF: Gaussian Avatar Reconstruction from Monocular Videos via Multi-view Diffusion” 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
Jiapeng Tang, Davide Davoli, Tobias Kirschstein, Liam Schoneveld, Matthias Niessner
GAF reconstructs photorealistic, animatable 3D Gaussian head avatars from short monocular videos by leveraging a specially trained multi-view diffusion model. The method effectively completes unobserved facial regions and maintains identity consistency, achieving superior fidelity in novel view and expression synthesis compared to prior approaches.
Key insights
- Specialized multi-view diffusion models, conditioned by pixel-aligned normal maps and input identity features, can robustly complete unobserved regions in monocular 3D head reconstruction.
- Iterative denoising to generate pseudo-ground truths for 3D supervision mitigates common issues like over-saturation and noisy gradients found in single-step Score Distillation Sampling.
- Distilling generative priors from 2D diffusion models into explicit 3D Gaussian Splatting representations offers a powerful paradigm for high-fidelity, real-time renderable avatars from limited input.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.