Depth Pro 2 CoreML Models

GitHub Technical Report Project Page All Sizes

Sharp, temporally consistent metric depth from a single camera, for photos and videos. These are the Core ML packages (.mlpackage) of the torch models on the Depth Pro 2 page.

Models

Larger models give sharper, steadier depth; smaller ones run faster. On a laptop, start with S.

  • Window. w4 is a temporal window of 4 frames. The model takes one frame at a time and keeps the earlier ones in its state, so create a new state for each new video.
  • Sizes. Each depth package has one function per input size, named res_<W>x<H>: res_854x480 (default), res_640x360, res_896x512, res_1138x640 and res_1280x720.
  • Field of view. The field-of-view package has a single size, 896x512, for 16:9 frames. It estimates the focal length of cameras that don't report it.
  • Memory. It grows quickly with the size: S needs about 1.5 GB at 854x480 and 5 GB at 1280x720; L about 4 GB and 18 GB. On a Mac with 16 GB, use S or B, or keep to the default size.

Requires Apple silicon and macOS 15 or later.

Try it

Clone the GitHub repository, put a depth package (and the field-of-view package, for cameras without a known focal length) next to each other, and run the Mac demo on it:

./demo/run.sh depthpro2-s_w4.mlpackage

To build a package at your own size -s HxW and window -w, see the demo's README:

uv run --extra coreml python -m demo.export stream -m depthpro2-s -w 4 -s 480x854

Use in Swift

import CoreML

let compiled = try MLModel.compileModel(at: URL(filePath: "depthpro2-s_w4.mlpackage"))
let config = MLModelConfiguration()
config.computeUnits = .cpuAndGPU
config.functionName = "res_854x480"  // omit for the default size
let model = try MLModel(contentsOf: compiled, configuration: config)

let state = model.makeState()  // a new state starts a new video

// frame: float16 MLMultiArray (1, 3, 480, 854), RGB in [0, 1]
let input = try MLDictionaryFeatureProvider(dictionary: ["frame": MLFeatureValue(multiArray: frame)])
let output = try model.prediction(from: input, using: state)
let inverseDepth = output.featureValue(for: "inverse_depth")!.multiArrayValue!  // (1, 480, 854)

inverse_depth is normalized by the focal length: depth in meters is f_px / (W * inverse_depth), with f_px the focal length in pixels of the H×W frame. The demo's Preprocess.swift shows how to prepare camera frames.

Videos and the torch models are on the Depth Pro 2 page.

License

The weights are released under the Apple Machine Learning Research Model license, for research use only: non-commercial scientific research and academic development.

Citation

@article{DepthPro2:2026:arxiv,
  title   = {Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth},
  author  = {Peiyun Hu and Marko Mihajlovic and Lyndon Duong and Lahav Lipson and Wei Dong and Ama\"{e}l Delaunoy and Stephan R. Richter and Vladlen Koltun},
  journal = {arXiv preprint arXiv:2611.26603},
  year    = {2026},
  url     = {https://arxiv.org/abs/2611.26603},
}
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including apple/DepthPro2-coreml

Paper for apple/DepthPro2-coreml