Reproduce “FFAvatar: Few-Shot, Feed-Forward, and Generalizable Avatar Reconstruction” 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
Thuan Hoang Nguyen, Jiahao Luo, Yinyu Nie, Hao Li, Gordon Guocheng Qian, Jian Wang
FFAvatar presents a feed-forward framework for reconstructing high-quality, animatable 3D Gaussian head avatars in seconds from a few unposed portrait images. This approach achieves higher PSNR, SSIM, LPIPS, and CSIM scores compared to previous state-of-the-art feed-forward methods on the NeRSemble benchmark.
Key insights
- A progressive training curriculum, combining broad-scale pretraining and targeted multi-view fine-tuning, is effective for achieving both strong generalization and high geometric fidelity.
- Learning FLAME parameters end-to-end directly from raw pixels eliminates the need for external preprocessing, enabling scalable training and faster inference for animating 3D Gaussian avatars.
- Fusing information from multiple unposed portrait images via a transformer-based Query-Former substantially enhances identity preservation and geometric consistency compared to single-view reconstruction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.