TeRA: Rethinking Text-guided Realistic 3D Avatar Generation

Yanwen Wang1*, Yiyu Zhuang1*, Jiawei Zhang1, Li Wang1, Yifei Zeng1, Xun Cao1, Xinxin Zuo2 Hao Zhu1✉
1Nanjing University, 2Concordia University
* Equal contribution, ✉ Corresponding author
Paper arXiv Video Code(Coming soon)
TeRA teaser

We propose TeRA, the first latent diffusion model specifically designed for text-guided 3D avatar generation. TeRA achieves superior inference speed, text-to-3D alignment, and visual quality, while naturally supporting text-guided structure-aware editing.

Abstract

Efficient 3D avatar creation is a significant demand in the metaverse, film/game, AR/VR, etc. 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.

Video

Method

Overall method. (a) Given the annotated multi-view human dataset, we train a text conditioned 3D avatar generative model. (b) The model is established upon a structured 3D human representation. The model training includes two stages: (c) firstly, a decoder is required by distilling a pretrained 3D human reconstruction model; (d) secondly, a structured latent diffusion model (LDM) is trained to generate structured latent maps from noises.

Results of Text-guided Generation

BibTeX

@article{wang2025TeRA,
  author    = {Wang, Yanwen and Zhuang, Yiyu and Zhang, Jiawei and Wang, Li and Zeng, Yifei and Cao, Xun and Zuo, Xinxin and Zhu, Hao},
  title     = {TeRA: Text-to-Avatar Generation via 3D Human Representation},
  journal   = {ICCV},
  year      = {2025},
}