Depth Anything 3: DA3METRIC-LARGE

# noqa: E501

Model Description

DA3 Metric Large model specialized for metric depth estimation in monocular settings, ideal for applications requiring real-world scale. Canonical metric depth; multiplying by focal length gives metric depth.

Capabilities

• ✅ Relative Depth
• ✅ Metric Depth
• ✅ Sky Segmentation

Quick Start

Installation

git clone https://github.com/ByteDance-Seed/depth-anything-3
cd depth-anything-3
pip install -e .

Basic Example

import torch
from depth_anything_3.api import DepthAnything3

# Load model from Hugging Face Hub
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DepthAnything3.from_pretrained("depth-anything/da3metric-large")
model = model.to(device=device)

# Run inference on images
images = ["image1.jpg", "image2.jpg"]  # List of image paths, PIL Images, or numpy arrays
prediction = model.inference(
    images,
    export_dir="output",
    export_format="glb"  # Options: glb, npz, ply, mini_npz, gs_ply, gs_video
)

# Access results
print(prediction.depth.shape)        # Depth maps: [N, H, W] float32
print(prediction.conf.shape)         # Confidence maps: [N, H, W] float32
print(prediction.extrinsics.shape)   # Camera poses (w2c): [N, 3, 4] float32
print(prediction.intrinsics.shape)   # Camera intrinsics: [N, 3, 3] float32

Command Line Interface

# Process images with auto mode
da3 auto path/to/images \
    --export-format glb \
    --export-dir output \
    --model-dir depth-anything/da3metric-large

# Use backend for faster repeated inference
da3 backend --model-dir depth-anything/da3metric-large
da3 auto path/to/images --export-format glb --use-backend

Model Details

• Developed by: ByteDance Seed Team
• Model Type: Vision Transformer for Visual Geometry
• Architecture: Plain transformer with unified depth-ray representation
• Training Data: Public academic datasets only

Key Insights

💎 A single plain transformer (e.g., vanilla DINO encoder) is sufficient as a backbone without architectural specialization. # noqa: E501

✨ A singular depth-ray representation obviates the need for complex multi-task learning.

Performance

🏆 Depth Anything 3 significantly outperforms:

• Depth Anything 2 for monocular depth estimation
• VGGT for multi-view depth estimation and pose estimation

For detailed benchmarks, please refer to our paper. # noqa: E501

Limitations

• The model is trained on academic datasets and may have limitations on certain domain-specific images # noqa: E501
• Performance may vary depending on image quality, lighting conditions, and scene complexity

Citation

If you find Depth Anything 3 useful in your research or projects, 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},  # noqa: E501
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2025}
}

Links

• 🏠 Project Page
• 📄 Paper
• 💻 GitHub Repository
• 🤗 Hugging Face Demo
• 📚 Documentation

Authors

Haotong Lin · Sili Chen · Junhao Liew · Donny Y. Chen · Zhenyu Li · Guang Shi · Jiashi Feng · Bingyi Kang # noqa: E501