DepthMaster: Taming Diffusion Models for Monocular Depth Estimation

Ziyang Song*, Zerong Wang*, Bo Li, Hao Zhang, Ruijie Zhu, Li Liu, Peng-Tao Jiang†, Tianzhu Zhang†,
*Equal Contribution, †Corresponding Author
University of Science and Technology of China, vivo Mobile Communication Co., Ltd.
Arxiv 2025

                       

We present DepthMaster, a tamed single-step diffusion model that customizes generative features in diffusion models to suit the discriminative depth estimation task. We introduce a Feature Alignment module to mitigate overfitting to texture and a Fourier Enhancement module to refine fine-grained details. DepthMaster exhibits state-of-the-art zero-shot performance and superior detail preservation ability, surpassing other diffusion-based methods across various datasets.

📢 News

2025-01-15: Evaluation code is released.
2025-01-02: Paper is released on arXiv.

Installation

Please refer to installation.md for installation.

Checkpoint

The model can be downloaded here.

🏃 Testing on your images

📷 Prepare images

Place your images in a directory, for example, under in_the_wild_example/input, and run the following inference command.

bash scripts/infer.sh

You can find all results in in_the_wild_example/output. Enjoy!

🦿 Evaluation on test datasets

Set data directory variable (also needed in evaluation scripts) and download evaluation datasets follow Marigold into corresponding subfolders:

export BASE_DATA_DIR=<YOUR_DATA_DIR>  # Set target data directory

wget -r -np -nH --cut-dirs=4 -R "index.html*" -P ${BASE_DATA_DIR} https://share.phys.ethz.ch/~pf/bingkedata/marigold/evaluation_dataset/

Download the model here to ckpt/eval subfolder. Run evaluation scripts, for example:

bash scripts/eval_kitti.sh

The evaluation results will be saved to output\kitti.

🎓 Citation

Please cite our paper:

@article{song2025depthmaster,
  title={DepthMaster: Taming Diffusion Models for Monocular Depth Estimation},
  author={Song, Ziyang and Wang, Zerong and Li, Bo and Zhang, Hao and Zhu, Ruijie and Liu, Li and Jiang, Peng-Tao and Zhang, Tianzhu},
  journal={arXiv preprint arXiv:2501.02576},
  year={2025}
}

Acknowledgements

The code is based on Marigold.

🎫 License

This work is licensed under the Apache License, Version 2.0 (as defined in the LICENSE).

By downloading and using the code and model you agree to the terms in the LICENSE.