Haodong Li^123§, Wangguangdong Zheng¹, Jing He³, Yuhao Liu¹, Xin Lin², Xin Yang³⁴,
Ying-Cong Chen^34✉, Chunchao Guo^1✉

¹Tencent Hunyuan ²UC San Diego ³HKUST(GZ) ⁴HKUST
^§Work primarily done during an internship at Tencent Hunyuan. ^✉Corresponding author.

DA² predicts dense, scale-invariant distance from a single 360° panorama in an end-to-end manner, with remarkable geometric fidelity and strong zero-shot generalization.

• We are currently uploading the curated panoramic data, please stay tuned!
• 2025-10-04 The 🤗Huggingface Gradio demo (online and local) are released!
• 2025-10-04 The inference code and the model are released!
• 2025-10-01 Paper released on arXiv!

This installation was tested on: Ubuntu 20.04 LTS, Python 3.12, CUDA 12.2, NVIDIA GeForce RTX 3090.

1. Clone the repository:

git clone https://github.com/EnVision-Research/DA-2.git
cd DA-2

2. Install dependencies using conda:

conda create -n da-2 python=3.12 -y
conda activate da-2
pip install -e src

For macOS users: Please remove xformers==0.0.28.post2 from src/pyproject.toml before pip install -e src, as xFormers does not support macOS.

1. Online demo: Hugggingface Space
2. Local demo:

python app.py

We've pre-uploaded the cases appeared in the project page. So you can proceed directly to step 3.

1. Images are placed in a directory, e.g., assets/demos.
2. (Optional) Masks (e.g., sky masks for outdoor images) in another directory, e.g., assets/masks. The filenames under both directories should be consistent.
3. Run the inference command:

sh infer.sh

4. The visualized distance and normal maps will be saved at output/infer/vis_all.png. The projected 3D point clouds will be saved at output/infer/3dpc.

If you find our work useful in your research, please consider citing our paper🌹:

@article{li2025depth,
  title={DA $\^{} 2$: Depth Anything in Any Direction},
  author={Li, Haodong and Zheng, Wangguangdong and He, Jing and Liu, Yuhao and Lin, Xin and Yang, Xin and Chen, Ying-Cong and Guo, Chunchao},
  journal={arXiv preprint arXiv:2509.26618},
  year={2025}
}

This implementation is impossible without the awesome contributions of MoGe, UniK3D, DINOv2, Accelerate, Gradio, HuggingFace Hub, and PyTorch to the open-cource community.