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 evaluation and 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(demo)	Dav2-large	335.3M	Download
Distill-Any-Depth-Dav2-Teacher-Large-2w-iter	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

Use from transformers

Here is how to use this model to perform zero-shot depth estimation:

from transformers import pipeline
from PIL import Image
import requests
# load pipe
pipe = pipeline(task="depth-estimation", model="xingyang1/Distill-Any-Depth-Large-hf")
# load image
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
# inference
depth = pipe(image)["depth"]

We are sincerely grateful to @keetrap and @Niels Rogge for their huge efforts in supporting our models in Transformers.

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}
}

Acknowledgements

Thanks to these great repositories: Depth Anything V2，MiDaS，GenPercept，GeoBench: 3D Geometry Estimation Made Easy，HDN，Detectron2 and many other inspiring works in the community.

Star History

License

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