Depth Pro

Sharp metric depth from a single photograph, with no camera intrinsics required, is what Apple's Depth Pro delivers: a 2.25 megapixel depth map with absolute scale in about 0.3 seconds on a standard GPU. The model also estimates focal length from the image itself, which is what allows metric output on arbitrary photos whose EXIF data is missing or wrong, and the paper emphasizes crisp boundary detail that prior monocular estimators tend to blur. Published at ICLR 2025, the repository provides a reference implementation whose retrained weights come close to, though do not exactly match, the numbers reported in the paper. Setup is a conda environment with a pip install of the package, a depth-pro-run command line tool, and a script that pulls the pretrained checkpoint, and inference runs on PyTorch, much faster with a CUDA GPU. Code and weights are released under Apple's own sample code license, which permits modification and redistribution with attribution but is not OSI approved, so commercial users should review the terms. Robotics, AR, and view synthesis projects adopted it as one of the strongest open monocular depth models.