Xiankang He1*,2 · Dongyan Guo1* · Hongji Li2,3
Ruibo Li4 · Ying Cui1 · Chi Zhang2✉
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 seizes are available in this repo.
- 2025-02-26:🔥🔥🔥 Paper, project page, code, models, and demos are released.
- Release training code.
- Release additional models in various sizes.
We provide two models of varying scales for robust relative depth estimation:
| Model | Architecture | Params | Checkpoint |
|---|---|---|---|
| Distill-Any-Depth-Multi-Teacher-Base | Dav2-base | 97.5M | Download |
| Distill-Any-Depth-Multi-Teacher-Large | Dav2-large | 335.3M | Download |
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:
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# 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 0 \ # Resolution for processing (0 to keep original resolution)
--output_dir output/${model} \ # Directory to save the output results
--arch_name 'depthanything-large' \ # Model architecture name (must match the pre-trained model)
done
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}
}This sample code is released under the MIT license. See LICENSE for more details.