---
pipeline_tag: unconditional-image-generation
license: apache-2.0
tags:
- 3D
- Gaussian Splatting
- GAN
- Human Head Synthesis
---

# CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis

CGS-GAN is a novel 3D Gaussian Splatting GAN framework designed for stable training and high-quality, 3D-consistent synthesis of human heads. It addresses the challenges of existing methods that compromise 3D consistency by relying on view-conditioning. CGS-GAN introduces a multi-view regularization technique to ensure training stability and a specialized generator architecture that facilitates efficient rendering and scaling, enabling output resolutions up to $2048^2$.

-   **Paper:** [CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis](https://huggingface.co/papers/2505.17590)
-   **Project Page:** [https://fraunhoferhhi.github.io/cgs-gan/](https://fraunhoferhhi.github.io/cgs-gan/)
-   **Repository:** [https://github.com/fraunhoferhhi/cgs-gan](https://github.com/fraunhoferhhi/cgs-gan)

## Usage

To run inference and render results, follow the setup instructions in the [official GitHub repository](https://github.com/fraunhoferhhi/cgs-gan) and use the `generate_samples.py` script with a pre-trained checkpoint:

```shell
python generate_samples.py --pkl path/to/network.pkl 
```

### Optional Parameters

-   `--truncation_psi`: Tradeoff between quality and variety (0: quality, 1: variety). Default is `0.8`.
-   `--num_ids`: Number of IDs to generate (number of rows). Default is `6`.
-   `--radius`: Radius of the camera. Default is `2.7`.
-   `--seed`: Random seed for generation. Default is `42`.
-   `--save_dir`: Directory to save the generated results. Default is `"results"`.

## Citation

Please cite our paper when using CGS-GAN in your work:

```bibtex
@misc{barthel2025cgsgan,
      title={CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis},
      author={Florian Barthel and Wieland Morgenstern and Paul Hinzer and Anna Hilsmann and Peter Eisert},
      year={2025},
      eprint={2505.17590},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.17590},
}
```