---
license: mit
library_name: pytorch
tags:
  - realdenseface
  - computer-vision
  - 3d-face-reconstruction
  - face-tracking
  - flame
  - dinov3
  - arxiv:2608.09238
---

# RealDenseFace

**Real-time Monocular 3D Face Reconstruction from Dense UV-space Priors**

Linzhou Li, Tianjia Shao, Kun Zhou  
State Key Lab of CAD&CG, Zhejiang University

[Project](https://gapszju.github.io/RealDenseFace/) · [Paper](https://arxiv.org/abs/2608.09238) · [Code](https://github.com/LinzhouLi/RealDenseFace)

RealDenseFace combines dense UV-space correspondence and relative-depth predictions with a tailored CUDA Gauss-Newton solver for FLAME reconstruction. It supports single-image fitting, offline sequence reconstruction, online tracking, and calibrated multi-view fitting.

## Models

| Model | Encoder | Input size | Checkpoint |
| --- | --- | --- | --- |
| ViT-S | DINOv3 ViT-S+ | 512 × 512 | [vits.pth](weights/realdenseface/vits.pth) |
| ViT-B | DINOv3 ViT-B | 512 × 512 | [vitb.pth](weights/realdenseface/vitb.pth) |

ViT-S is the default model for online tracking; ViT-B is the default for single-image and offline reconstruction. Matching model and fitting configurations are included in `configs/`.

Additional inference resources are included: [FaceBox weights](weights/facebox/face_box.pth) and [FLAME assets](weights/flame/flame_assets.npz).

For installation and usage, see the [GitHub repository](https://github.com/LinzhouLi/RealDenseFace). The FLAME model `flame2023.pkl` must be obtained separately from the [official website](https://flame.is.tue.mpg.de/) under its license. Third-party resources retain their respective licenses.

## Citation

```bibtex
@article{li2026realdenseface,
  title   = {RealDenseFace: Real-time Monocular 3D Face Reconstruction from Dense UV-space Priors},
  author  = {Li, Linzhou and Shao, Tianjia and Zhou, Kun},
  journal = {arXiv preprint arXiv:2608.09238},
  year    = {2026}
}
```