TAGA: Self-supervised Learning for Template-free Animatable Gaussian Avatars
Keywords: Template-free avatar, Animatble Avatar, Gaussian Splatting, Self-supervised Learning
Abstract: Decoupling from customized parametric templates marks an integral leap towards creating fully flexible, animatable avatars. In this work, we introduce TAGA (Template-free Animatable Gaussian Avatars), the first template-free, Gaussian-based solution for the reconstruction of animatable avatars from monocular videos, which offers distinct advantages in fast training and real-time rendering. Constructing template-free avatars is challenging due to the lack of predefined shapes and reliable skinning anchors to ensure consistent geometry and movement. TAGA addresses this by introducing a self-supervised method which guides both geometry and skinning learning leveraging the one-to-one correspondence between canonical and observation spaces. During the forward mapping phase, a voxel-based skinning field is introduced to learn smooth deformations that generalize to unseen poses. However, without template priors, forward mapping often captures spurious correlations of adjacent body parts, leading to unrealistic geometric artifacts in the canonical pose. To alleviate this, we define Gaussians with spurious correlations as "Ambiguous Gaussians'' and then propose a new backward mapping strategy that integrates anomaly detection to identify and correct Ambiguous Gaussians. Compared to existing state-of-the-art template-free methods, TAGA achieves superior visual fidelity for novel views and poses, while being 60 $\times$ faster in training (0.5 hours vs 30 hours) and 560 $\times$ faster in rendering (140 FPS vs 0.25 FPS). Experiments on challenging datasets that possess limited pose diversity further demonstrate TAGA’s robustness and generality. Code will be released.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2009
Loading