Distill Any Depth: Distillation Creates a Stronger Monocular Depth Estimator

Xiankang He^1*,2 · Dongyan Guo^1* · Hongji Li^2,3
Ruibo Li⁴ · Ying Cui¹ · Chi Zhang^2✉

¹ZJUT   ²WestLake University   ³LZU   ⁴NTU
✉ Corresponding author
*Equal Contribution. This work was done while Xiankang He was visiting Westlake University.

We present Distill-Any-Depth, a new SOTA monocular depth estimation model trained with our proposed knowledge distillation algorithms. Models with various sizes are available in this repo.

News

• 2025-03-08: We release the small size of our model(Dav2).
• 2025-03-02:🔥🔥🔥 Our demo is updated to GPU version. Enjoy it! We also include the Gradio demo code in this repo.
• 2025-02-26:🔥🔥🔥 Paper, project page, code, models, and demos are released.

TODO

• Release training code.
• Release additional models in various sizes.

Pre-trained Models

We provide two models of varying scales for robust relative depth estimation:

Model	Architecture	Params	Checkpoint
Distill-Any-Depth-Multi-Teacher-Small	Dav2-small	24.8M	Download
Distill-Any-Depth-Multi-Teacher-Base	Dav2-base	97.5M	Download
Distill-Any-Depth-Multi-Teacher-Large	Dav2-large	335.3M	Download

Getting Started

We recommend setting up a virtual environment to ensure package compatibility. You can use miniconda to set up the environment. The following steps show how to create and activate the environment, and install dependencies:

# Create a new conda environment with Python 3.10
conda create -n distill-any-depth -y python=3.10

# Activate the created environment
conda activate distill-any-depth

# Install the required Python packages
pip install -r requirements.txt

# Navigate to the Detectron2 directory and install it
cd detectron2
pip install -e .

cd ..
pip install -e .

To download pre-trained checkpoints follow the code snippet below:

Running from commandline

We provide a helper script to run the model on a single image directly:

# Run prediction on a single image using the helper script
source scripts/00_infer.sh
# or use bash
bash scripts/00_infer.sh

# you should download the pretrained model and input the path on the '--checkpoint'

# Define the GPU ID and models you wish to run
GPU_ID=0
model_list=('xxx')  # List of models you want to test

# Loop through each model and run inference
for model in "${model_list[@]}"; do
    # Run the model inference with specified parameters
    CUDA_VISIBLE_DEVICES=${GPU_ID} \
    python tools/testers/infer.py \
        --seed 1234 \  # Set random seed for reproducibility
        --checkpoint 'checkpoint/large/model.safetensors' \  # Path to the pre-trained model checkpoint
        --processing_res 700 \ 
        --output_dir output/${model} \  # Directory to save the output results
        --arch_name 'depthanything-large' \  # [depthanything-large, depthanything-base]
done

Gradio demo

We also include the Gradio demo code, Please clone the project and set up the environment using pip install.

# Create a new conda environment with Python 3.10
conda create -n distill-any-depth -y python=3.10

# Activate the created environment
conda activate distill-any-depth

# Install the required Python packages
pip install -r requirements.txt

pip install -e .

Make sure you can connect to Hugging Face, or use the local path. (app.py)

# if use hf_hub_download, you can use the following code
checkpoint_path = hf_hub_download(repo_id=f"xingyang1/Distill-Any-Depth", filename=f"large/model.safetensors", repo_type="model")

# if use local path, you can use the following code
# checkpoint_path = "path/to/your/model.safetensors"

in the end,

python app.py

:~/Distill-Any-Depth-main# python app.py 
xFormers not available
xFormers not available
xFormers not available
xFormers not available
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
IMPORTANT: You are using gradio version 4.36.0, however version 4.44.1 is available, please upgrade.
--------

More Results

Citation

If you find our work useful, please cite the following paper:

@article{he2025distill,
  title   = {Distill Any Depth: Distillation Creates a Stronger Monocular Depth Estimator},
  author  = {Xiankang He and Dongyan Guo and Hongji Li and Ruibo Li and Ying Cui and Chi Zhang},
  year    = {2025},
  journal = {arXiv preprint arXiv: 2502.19204}
}

Related Projects

• Depth Anything V2
• MiDaS
• GenPercept
• GeoBench: 3D Geometry Estimation Made Easy
• HDN
• Detectron2

Star History

License

This sample code is released under the MIT license. See LICENSE for more details.