UniLat3D: Geometry-Appearance Unified Latents for Single-Stage 3D Generation

TL;DR

UniLat3D encodes shape and appearance together in one compact latent representation, allowing a single flow-matching model to generate textured 3D assets directly in one stage.

Abstract

Many pretrained 3D generators use a two-stage pipeline that first produces geometry and then synthesizes appearance. This separation can cause geometry-texture misalignment and adds substantial computation. UniLat3D introduces a unified framework that encodes both properties in a single latent space for direct, single-stage generation. Its geometry-appearance Unified VAE compresses high-resolution sparse features into UniLat, a compact dense representation that combines structural and visual information and can be decoded into formats such as 3D Gaussians and meshes. A single flow-matching model then maps Gaussian noise directly to UniLat, eliminating redundant generation stages. Trained only on public datasets, UniLat3D produces detailed 3D assets from one image in seconds and achieves strong geometry and appearance quality.

Key contributions

• Unifies geometry and appearance in one compact 3D latent representation.
• Generates textured assets with a single flow-matching model instead of a two-stage pipeline.
• Decodes the shared representation into multiple 3D formats, including Gaussians and meshes.

Citation

@inproceedings{wu2026unilat3d,
  title={UniLat3D: Geometry-Appearance Unified Latents for Single-Stage 3D Generation},
  author={Wu, Guanjun and Fang, Jiemin and Yang, Chen and Li, Sikuang and Yi, Taoran and Lu, Jia and Zhou, Zanwei and Cen, Jiazhong and Xie, Lingxi and Zhang, Xiaopeng and Wei, Wei and Liu, Wenyu and Wang, Xinggang and Tian, Qi},
  booktitle={Findings of CVPR},
  year={2026},
  doi={10.48550/arXiv.2509.25079}
}