Apple's machine learning group has published Depth Pro 2, a model that estimates how far away each pixel of an image is, in meters, using a single ordinary camera. The page is dated November 2026 and lists code and model weights as released.
What the model does
One network covers three inputs: a single photo, a complete video and a live camera feed. It works at the full resolution of whatever it is given. The authors say its output stays sharp at object edges and steady from one frame to the next.
The second point is the claim behind the paper. According to the authors, earlier monocular depth methods forced a choice. Approaches that look at several frames give stable results over time, while single-frame approaches give crisper outlines. Depth Pro 2 is presented as delivering accurate metric depth, sharp boundaries and temporal consistency together, ahead of prior work on each of those measures.
What is released
Apple publishes a family of models rather than one checkpoint. The authors say the family sets a new Pareto frontier across metric accuracy, sharpness, temporal consistency and runtime, meaning teams can pick a variant that suits their speed needs. Code is on GitHub, and weights are available through the same page.
The page also links two earlier Apple releases: the original Depth Pro from October 2024, and SHARP, a single-image view synthesis method from December 2025.
What the source does not say
The page gives no licence terms, hardware requirements or runtime figures for the new family, and no product or service is attached. All performance claims come from the authors. Check the repository for licence and usage terms before any commercial use.
Who should look
If your processes involve judging size or distance from images, this is a free model to benchmark against your own photos. If they do not, there is nothing to act on.