mrgnw/depth-anything-v2-coreml

Depth Anything V2 - CoreML

Depth Anything V2 models (Base and Large) converted to CoreML format for optimized inference on Apple Silicon (M-series chips).

Models

Model	Size	Parameters	Performance (M4 Pro est.)	License
Small F16	48 MB	24.8M	30ms (33 fps)	Apache-2.0
Base F16	172 MB	97.5M	60-90ms (14 fps)	CC-BY-NC-4.0
Large F16	590 MB	335.3M	200-300ms (4 fps)	CC-BY-NC-4.0

All models use Float16 precision and run on Apple's Neural Engine + GPU + CPU.

License

Both Base and Large models are CC-BY-NC-4.0 (non-commercial only), following the official Depth Anything V2 licensing.

For commercial use, you must use the Small model (Apache-2.0), which is available directly from Apple's CoreML model zoo.

Download

Base model:

curl -L -o DepthAnythingV2BaseF16.mlpackage.tar.gz \
  "https://huggingface.co/mrgnw/depth-anything-v2-coreml/resolve/main/DepthAnythingV2BaseF16.mlpackage.tar.gz"
tar -xzf DepthAnythingV2BaseF16.mlpackage.tar.gz

Large model:

curl -L -o DepthAnythingV2LargeF16.mlpackage.tar.gz \
  "https://huggingface.co/mrgnw/depth-anything-v2-coreml/resolve/main/DepthAnythingV2LargeF16.mlpackage.tar.gz"
tar -xzf DepthAnythingV2LargeF16.mlpackage.tar.gz

Small model (from Apple):

curl -L -o DepthAnythingV2SmallF16.mlpackage.zip \
  "https://ml-assets.apple.com/coreml/models/Image/DepthEstimation/DepthAnything/DepthAnythingV2SmallF16.mlpackage.zip"
unzip DepthAnythingV2SmallF16.mlpackage.zip

Usage

Swift

import CoreML

let modelURL = URL(fileURLWithPath: "DepthAnythingV2BaseF16.mlpackage")
let config = MLModelConfiguration()
config.computeUnits = .all  // Use Neural Engine + GPU + CPU

let model = try MLModel(contentsOf: modelURL, configuration: config)
// Input: RGB image (1, 3, 518, 518)
// Output: depth map (1, 518, 518)

Performance

M4 Pro (estimated):

• Small: ~25-30ms per frame
• Base: ~60-90ms per frame
• Large: ~200-300ms per frame

These are 10-20x faster than ONNX CPU inference because they use the Apple Neural Engine.

Citation

@article{yang2024depth,
  title={Depth Anything V2},
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  journal={arXiv:2406.09414},
  year={2024}
}

Related

• Original Depth Anything V2
• spatial-maker - Uses these models for spatial video/photo conversion