What Apple actually released
Depth Pro 2 estimates how far every pixel in an image is from the camera, in meters, using one ordinary camera. That is two hard problems in one sentence. Monocular means there is no second lens and no lidar to measure against, so distance has to be inferred from perspective, texture and the apparent size of familiar things. Metric means the answer comes back in real units rather than a ranking of nearer and farther, which is what an AR placement, a measurement or a robot needs from the output, instead of just a blur effect.
The release is a family, not a single checkpoint. Four sizes run from the 24M-parameter S model, a 97 MB download the README suggests for laptops, up to the 872M-parameter H+, at 3.5 GB, which is the default and the one used for the project page's native-resolution comparisons. All four share a field-of-view estimator that predicts the focal length when the file carries no usable camera metadata; without a focal length, pixel geometry does not convert into meters.
The design change from the first Depth Pro is time. Version one worked on single frames. Version two processes frames jointly: a multi-view transformer shares information across frames, alternating attention within a frame and across frames after its first third of blocks. The point of that machinery is a familiar failure mode: run a still-image model frame by frame and the depth flickers, because each frame is a fresh guess.
The claim, and what is printed underneath it
Apple's pitch is that the usual trade-off is gone. Multi-frame methods tend to be steadier and softer at edges; single-frame methods tend to be sharper and less stable. Depth Pro 2, Apple says, is metrically accurate, sharp and temporally consistent at once, and outperforms prior single-frame and multi-frame work on Boundary F1 and Jitter, averaged over the Sintel and Spring datasets.
Read the project page closely and the support is thinner than the sentence. The page states the win and prints no scores, no table and no named baseline figures. The comparison videos carry their own caveat in fine print: Depth Pro 2 H+ runs at native resolution while baselines run at their default settings, which is a reasonable way to demo sharpness and not a controlled comparison.
The paper does contain numbers, but its headline device is average rank across datasets, which orders methods without telling you the size of any gap. The most concrete figure, from alphaXiv's summary of the paper's tables, is a real one: on Sintel video with predicted camera intrinsics, H+ puts 41.4% of pixels within a factor of 1.25 of the true depth, against 37.0% for UniDepthV2-L. That is a genuine lead on a hard dataset, and it is also a reminder of the state of the field: on this test, roughly six in ten pixels from the best model Apple has still fall outside that factor-of-1.25 band.
Speed, Apple does publish, on the model cards, for the largest model on a datacenter GPU. On an NVIDIA B200, in bfloat16, processing a 32-frame clip at native resolution, H+ runs at 95.9 frames per second at VGA, 20.2 at 720p, 5.3 at 1080p and 0.4 at 4K, with peak memory rising from 8.3 GiB at VGA to 114.6 GiB at 4K. Those figures say nothing about live speed on a laptop or a phone, or about the smaller models. What the project page and the README do not publish is any comparison with the original Depth Pro, so existing users cannot tell from the release what the upgrade buys them.
The license is the product decision
The weights are public. What you may do with them is narrow. They ship under the Apple Machine Learning Research Model license (the repo's LICENSE_MODEL file), permitting use exclusively for research purposes, defined as non-commercial research and academic activities. The code is on GitHub under Apple's standard software license; the weights are the valuable part, and the weights are the restricted part. There is no training code and no evaluation code, so the published benchmark results cannot be rerun from what Apple shipped, though the inference code and the weights do let you test the models on your own footage, subject to the license.
For a research lab, none of that is a problem: the family is a strong, inspectable baseline, and the small S model makes it easy to start. For anyone building a product, the release is a demonstration, not an input. You can prototype a video effect, an AR sketch or a robotics experiment to learn whether metric depth helps your pipeline. The moment the answer is yes, you need a different model, because these weights cannot go into the product.
The practical notes, if you try it
Start with S on your own footage and judge edges and flicker there, not in Apple's gallery; step up a size only if the boundaries look wrong. Fixed-lens footage suits the release best, because a video gets one focal length for all its frames, and a zoom works against that assumption. And read the card's B200 figures as a starting point, not a promise: live use on your own hardware, at your resolution, is a measurement you have to make.
Sources
- [1] Apple, Depth Pro 2 project pageRead source
- [2] Hu et al., "Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth", arXiv:2611.26603Read source
- [3] Apple, ml-depth-pro-2 repository (README and LICENSE_MODEL), GitHubRead source
- [4] Apple, DepthPro2-H model card, Hugging FaceRead source
- [5] alphaXiv, summary of arXiv:2611.26603Read source