StereoPilot: Learning Unified and Efficient Stereo Conversion via Generative Priors

Guibao Shen^1,3†, Yihua Du¹, Wenhang Ge^1,3*†, Jing He¹, Chirui Chang³, Donghao Zhou⁴, Zhen Yang¹, Luozhou Wang¹, Xin Tao³, Ying-Cong Chen^1,2‡**

¹HKUST(GZ), ²HKUST, ³Kling Team, Kuaishou Technology, ⁴CUHK

(*Equal contribution, †This work was conducted during the author's internship at Kling, ‡Corresponding author)

📖 Introduction

TL;DR: We propose StereoPilot, an efficient feed-forward architecture that leverages pretrained video diffusion transformers to directly synthesize novel views, overcoming the limitations of Depth-Warp-Inpaint methods without iterative denoising. With a domain switcher and cycle consistency loss, it enables robust multi-format stereo conversion. We also introduce UniStereo, the first large-scale unified dataset featuring both parallel and converged stereo formats.

🎬 Click the image to view our showcase video

🔥 Updates

• [2025.12.16]: Release inference code and Project Page.

⚙️ Requirements

Our inference environment:

• Python 3.12
• CUDA 12.1
• PyTorch 2.4.1
• GPU: NVIDIA A800 (only ~23GB VRAM required)

🛠️ Installation

Step 1: Clone the repository

git clone https://github.com/KlingTeam/StereoPilot.git

cd StereoPilot

Step 2: Create conda environment

conda create -n StereoPilot python=3.12

conda activate StereoPilot

Step 3: Install dependencies

pip install -r requirements.txt

pip install flash-attn==2.7.4.post1 --no-build-isolation

Step 4: Download model checkpoints

Place the following files in the ckpt/ directory:

Download StereoPilot.safetensor & Wan2.1-1.3B base model:

pip install "huggingface_hub[cli]"

huggingface-cli download KlingTeam/StereoPilot StereoPilot.safetensors --local-dir ./ckpt

huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir ./ckpt/Wan2.1-T2V-1.3B

🚀 Inference

Input Requirements

For each input video, you need:

1. Video file (.mp4): Monocular video, 81 frames, 832×480 resolution, 16fps
2. Prompt file (.txt): Text description of the video content (same name as video)

Example (you can try the cases in the sample/ folder):

sample/
├── my_video.mp4
└── my_video.txt   

Running Inference

Basic usage:

# Edit toml/infer.toml to customize model paths. If you followed the above steps, there is no need to change
python sample.py \
  --config toml/infer.toml \
  --input /path/to/input_video.mp4 \
  --output_folder /path/to/output \
  --device cuda:0

Using the example script:

bash sample.sh

Generate Stereo Visualization

After inference, you can generate Side-by-Side (SBS) and Red-Cyan anaglyph stereo videos for visualization:

python utils/stereo_video.py \
  --left /path/to/left_eye.mp4 \
  --right /path/to/right_eye.mp4 \

Output files:

📊 Dataset

We introduce UniStereo, the first large-scale unified stereo video dataset featuring both parallel and converged stereo formats.

UniStereo consists of two parts:

• 3DMovie - Converged stereo format from 3D movies
• Stereo4D - Parallel stereo format (coming soon)

For detailed data processing instructions, please refer to StereoPilot_Dataprocess.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Wan2.1 - Base video generation model
• Diffusion Pipe - Training code base

🌟 Citation

If you find our work helpful, please consider citing:

@misc{shen2025stereopilot,
  title={StereoPilot: Learning Unified and Efficient Stereo Conversion via Generative Priors},
  author={Shen, Guibao and Du, Yihua and Ge, Wenhang and He, Jing and Chang, Chirui and Zhou, Donghao and Yang, Zhen and Wang, Luozhou and Tao, Xin and Chen, Ying-Cong},
  year={2025},
  eprint={2512.16915},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.16915}, 
}