Depth Anything 3 ONNX

ONNX-compatible implementation of Depth Anything 3 for efficient depth estimation inference.

This repository provides scripts to export Depth Anything 3 models to ONNX format and run inference using ONNXRuntime, enabling deployment on a wide range of platforms with minimal dependencies.

Get Started in 5 minutes →

Features

• ONNX Export: Convert Depth Anything 3 PyTorch models to ONNX format
• Minimal Dependencies: Inference requires only ONNXRuntime, OpenCV, and NumPy
• Multiple Model Sizes: Support for DA3-Small, DA3-Base, DA3-Large models
• Dynamic Input Support: Flexible input resolution handling
• Easy Integration: Simple API for depth estimation in your projects

Performance

ONNX models provide significant performance benefits:

• Reduced Memory: No PyTorch runtime overhead
• Cross-Platform: Runs on CPU, CUDA, TensorRT, DirectML, Coreml, etc.
• Optimized: ONNXRuntime automatically optimizes the graph
• Depth Anything 3 ONNX Performance Report
  • Model: depth-anything/DA3-SMALL
  • Platform: macOS M4 (CPU)
  • Runtime: ONNX Runtime with CPUExecutionProvider

Installation

Requirements

pip install -r requirements.txt

For Export Only

If you only need to run inference with pre-exported ONNX models:

pip install onnxruntime opencv-python pillow imageio numpy matplotlib

For Export

To export PyTorch models to ONNX, you'll need PyTorch:

pip install torch>=2.0.0 torchvision
pip install -r requirements.txt

Quick Start

1. Export Model to ONNX

Export a Depth Anything 3 model from HuggingFace to ONNX format:

python export_onnx.py \
    --model depth-anything/DA3-SMALL \
    --process-res 504 \
    --output DA3-SMALL-504.onnx

By default the exported model supports dynamic batch, height, and width axes, and the script runs portable ONNX Runtime basic graph optimization.

Available Models:

• depth-anything/DA3-SMALL (80M params)
• depth-anything/DA3-BASE (120M params)
• depth-anything/DA3-LARGE (350M params)

Note: The first run will download the model from HuggingFace. Models are cached in ~/.cache/huggingface/.

For mobile converters that prefer a smaller fixed-shape graph, export a static model:

python export_onnx.py \
    --model depth-anything/DA3-SMALL \
    --process-res 504 \
    --output DA3-SMALL-504-static.onnx \
    --static

2. Run Inference

Run depth estimation on an image using the exported ONNX model:

python run_onnx.py \
    --image assets/examples/input.jpg \
    --onnx DA3-SMALL-504.onnx \
    --process-res 504 \
    --export-dir ./output \
    --visualize

Arguments:

• --image: Path to input image
• --onnx: Path to ONNX model file
• --process-res: Processing resolution (default: 504)
• --export-dir: Directory to save depth visualization
• --visualize: Display visualization with matplotlib

3. Use in Your Code

import cv2
import numpy as np
import onnxruntime as ort
from PIL import Image

# Load ONNX model
sess = ort.InferenceSession("DA3-SMALL-504.onnx")

# Load and preprocess image
image = cv2.imread("input.jpg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (504, 504))
image = image.transpose(2, 0, 1).astype(np.float32) / 255.0
image = np.expand_dims(image, axis=0)

# Run inference
depth = sess.run(None, {"image": image})[0]

# Depth shape: (1, 1, 504, 504)
print(f"Depth map shape: {depth.shape}")

Model Export Details

The export script creates ONNX models with:

• Input: image tensor of shape (batch, 3, height, width), dtype float32, range [0, 1]
• Output: depth tensor of shape (batch, 1, height, width), dtype float32
• Dynamic Axes: Batch size, height, and width are dynamic unless --static is used
• Input Size Constraint: Height and width must be divisible by the DA3 patch size (14). run_onnx.py handles this through preprocessing.
• ONNX Opset: 17 (configurable via --opset)
• Graph Optimization: ONNX Runtime basic optimization is enabled by default

Export Options

python export_onnx.py --help

Key Parameters:

• --model: HuggingFace model ID or local path
• --process-res: Fixed square resolution for export (e.g., 504, 672)
• --output: Output ONNX file path
• --opset: ONNX opset version (default: 17)
• --static: Export fixed batch/height/width axes for mobile-friendly conversion
• --optimize: ONNX Runtime graph optimization level: none, basic, extended, or all (default: basic)

Inference Details

The inference script handles:

1. Image loading and preprocessing
2. Fixed-square resizing to the requested processing resolution
3. ONNX model inference
4. Depth map post-processing and visualization
5. Resizing output back to original image resolution

Dynamic ONNX exports can be used with different --process-res values at inference time. Static exports must be run at the same resolution used during export.

Inference Options

python run_onnx.py --help

Recommended Execution Providers

import onnxruntime as ort

# For CUDA GPU
sess = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

# For CPU
sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])

# For TensorRT (fastest on NVIDIA GPUs)
sess = ort.InferenceSession("model.onnx", providers=["TensorrtExecutionProvider"])

Project Structure

Depth-Anything-3-Onnx/
├── src/
│   └── depth_anything_3/       # Source code from Depth Anything 3
├── assets/
│   └── examples/               # Sample images
├── export_onnx.py              # Export PyTorch to ONNX
├── run_onnx.py                 # Run inference with ONNX
├── requirements.txt            # Python dependencies
├── LICENSE                     # Apache 2.0 License
└── README.md                   # This file

Troubleshooting

Issue: "No module named 'depth_anything_3'"

Make sure you're running scripts from the project root directory where the src/ folder is located.

Issue: Model download fails

If you experience network issues downloading from HuggingFace:

export HF_ENDPOINT=https://hf-mirror.com

Issue: ONNX inference is slow on CPU

Try using a smaller model (DA3-SMALL) or enabling optimizations:

sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
sess = ort.InferenceSession("model.onnx", sess_options=sess_options)

Citation

If you use Depth Anything 3 in your research, please cite:

@article{depthanything3,
  title={Depth Anything 3: Recovering the visual space from any views},
  author={Haotong Lin and Sili Chen and Jun Hao Liew and Donny Y. Chen and Zhenyu Li and Guang Shi and Jiashi Feng and Bingyi Kang},
  journal={arXiv preprint arXiv:2511.10647},
  year={2025}
}

Acknowledgments

• Depth Anything 3: Original Repository
• Depth Anything ONNX: Reference Implementation for Depth Anything V1/V2
• Thanks to the Depth Anything team for their amazing work on monocular depth estimation

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

The Depth Anything 3 models have varying licenses:

• Apache 2.0: DA3-SMALL, DA3-BASE, DA3METRIC-LARGE, DA3MONO-LARGE
• CC BY-NC 4.0: DA3-GIANT, DA3-LARGE, DA3NESTED-GIANT-LARGE

Please refer to the official Depth Anything 3 repository for model-specific license information.

Related Projects

• Depth Anything 3 - Official PyTorch implementation
• Depth Anything V2 - Previous version
• Depth Anything ONNX - ONNX implementation for V1/V2

Contributing

Contributions are welcome! Please feel free to submit issues or pull requests.