Depth Pro — MLX

Apple Depth Pro (sharp monocular metric depth from a single image) converted to MLX for Apple Silicon.

Code, CLI and 3D photo pipeline: github.com/ntoantoan/depthpro-mlx

Usage

pip install "depthpro-mlx[inpaint]"
depthpro-mlx photo.jpg --ply     # depth.png, pointcloud.ply, 3d.mp4

import numpy as np
from PIL import Image
from depthpro_mlx import load_model, predict

model = load_model()                      # downloads these weights
image = np.array(Image.open("photo.jpg").convert("RGB"))
out = predict(model, image)               # f_px=... if you know the focal length
out["depth"]                              # (H, W) depth in meters
out["focallength_px"]                     # estimated focal length

Files

File	Description
`depth_pro.safetensors`	All 952M parameters in float16, NHWC conv layout

Accuracy and speed

Parity against the official PyTorch implementation on the reference image (3024×2268):

	Time (M4 Pro)	Peak memory	AbsRel vs. PyTorch fp32
PyTorch MPS fp16	3.76 s	7.0 GB	–
MLX fp16	3.06 s	7.5 GB	6.2e-4
MLX fp32	4.17 s	15.0 GB	1.7e-6

License

These weights are a format conversion of Apple's released Depth Pro weights and are redistributed under Apple's original license terms (see LICENSE). That license is not OSI-approved: Apple grants a personal, non-exclusive license to use, modify and redistribute, and does not explicitly address commercial use, so check the terms before shipping the weights in a product. This conversion is not affiliated with or endorsed by Apple.

@inproceedings{Bochkovskii2024:arxiv,
  author    = {Aleksei Bochkovskii and Ama\"{e}l Delaunoy and Hugo Germain and Marcel Santos and
               Yichao Zhou and Stephan R. Richter and Vladlen Koltun},
  title     = {Depth Pro: Sharp Monocular Metric Depth in Less Than a Second},
  booktitle = {International Conference on Learning Representations},
  year      = {2025},
  url       = {https://arxiv.org/abs/2410.02073},
}