In this paper, we rethink text-to-avatar generative models by proposing TeRA, a more efficient and effective framework than the previous SDS-based models and general large 3D generative models. Our approach employs a two-stage training strategy for learning a native 3D avatar generative model. Initially, we distill a decoder to derive a structured latent space from a large human reconstruction model. Subsequently, a text-controlled latent diffusion model is trained to generate photorealistic 3D human avatars within this latent space. TeRA enhances the model performance by eliminating slow iterative optimization and enables text-based partial customization through a structured 3D human representation. Experiments have proven our approach's superiority over previous text-to-avatar generative models in subjective and objective evaluation.
@article{arxiv.2509.02466,
title = {TeRA: Rethinking Text-guided Realistic 3D Avatar Generation},
author = {Yanwen Wang and Yiyu Zhuang and Jiawei Zhang and Li Wang and Yifei Zeng and Xun Cao and Xinxin Zuo and Hao Zhu},
journal= {arXiv preprint arXiv:2509.02466},
year = {2025}
}