D³-Predictor: Noise-Free Deterministic Diffusion for Dense Prediction

[🌐 Website] • [📜 Paper] • [🤗 HF Models] • [🤗 HF Dataset] •

Repo for "D³-Predictor: Noise-Free Deterministic Diffusion for Dense Prediction"

D³-Predictor Model

News

• 2025.12: We release the [Training data] .

• 2025.12: We release the checkpoints of the [D³-Predictor model] .

Installation

Create a virtual environment and install the required dependencies:

conda create -n d3_predictor python=3.10
conda activate d3_predictor
pip install -r requirements.txt

Training

Please follow the steps below for training:

• For the depth estimation task, use the code in D3-Predictor-Depth/.
• For the surface normal estimation task, use the code in D3-Predictor-Normal/.
• For the image matting task, use the code in D3-Predictor-Matting/.

1. Make sure to set the correct paths in .yaml configs.
2. Specify the wandb API key before training.
3. Start training

bash launch_train_{depth/normal/matting}.sh

Inference

1. Download checkpoints to checkpoints/{depth/normal/matting}
2. Make sure to set the correct paths in inference_{depth/normal/matting}.py.
3. Start inference

python D3-Predictor-{Depth/Normal/Matting}/inference_{depth/normal/matting}.py

Acknowledgment

Thanks to Marigold for data preprocessing and results evaluation support, Stable Diffusion 2.1 and FLUX.1-dev for powerful pretrained model, and Cleandift for their wonderful open-sourced work.

Citation

If you find it helpful, please kindly cite the paper.

@misc{xia2025mathrmdmathrm3predictornoisefreedeterministicdiffusion,
      title={D³-Predictor: Noise-Free Deterministic Diffusion for Dense Prediction}, 
      author={Changliang Xia and Chengyou Jia and Minnan Luo and Zhuohang Dang and Xin Shen and Bowen Ping},
      year={2025},
      eprint={2512.07062},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.07062}, 
}

📬 Contact

If you have any inquiries, suggestions, or wish to contact us for any reason, we warmly invite you to email us at 202066@stu.xjtu.edu.cn or cp3jia@stu.xjtu.edu.cn.