gym890/diffportrait360

arxiv.org/abs/2503.15667

[CVPR'25]DiffPortrait360: Consistent Portrait Diffusion for 360 View Synthesis

Yuming Gu^1,2 · Phong Tran² · Yujian Zheng² · Hongyi Xu³ · Heyuan Li⁴ · Adilbek Karmanov² · Hao Li^2,5
1Unviersity of Southern California  ²MBZUAI   ³ByteDance Inc.  
⁴The Chinese University of Hong Kong, Shenzhen  ⁵Pinscreen Inc.

📜 Requirements

• An NVIDIA GPU with CUDA support is required.
  • We have tested on a single A6000 GPU.
  • Minimum: The minimum GPU memory required is 30GB for generating a single NVS video (batch_size=1) of 32 frames each time.
  • Recommended: We recommend using a GPU with 40GB of memory.
• Operating system: Linux

🧱 Download Pretrained Models

Diffportrait360
|----...
|----pretrained_weights
  |----back_head-230000.th # back head generator
  |----model_state-3400000.th # diffportrait360 main module
  |----easy-khair-180-gpc0.8-trans10-025000.th
|----...

🔗 BibTeX

If you find Diffportrait360 is useful for your research and applications, please cite Diffportrait360 using this BibTeX:

@article{gu2025diffportrait360,
  title={DiffPortrait360: Consistent Portrait Diffusion for 360 View Synthesis},
  author={Gu, Yuming and Tran, Phong and Zheng, Yujian and Xu, Hongyi and Li, Heyuan and Karmanov, Adilbek and Li, Hao},
  journal={arXiv preprint arXiv:2503.15667},
  year={2025}
}

License

Our code is distributed under the Apache-2.0 license.

Acknowledgements

This work is supported by the Metaverse Center Grant from the MBZUAI Research Office. We appreciate the contributions from Diffportrait3D, PanoHead, SphereHead, ControlNet for their open-sourced research. We thank Egor Zakharov, Zhenhui Lin, Maksat Kengeskanov, and Yiming Chen for the early discussions, helpful suggestions, and feedback.