News

Apple presents Depth Pro 2 for images and video

Depth Pro 2 estimates depth in meters for every pixel at the original resolution. Apple provides inference code and research-use weights.

D
Nov 20, 2026 · 2 min read

Apple researchers presented Depth Pro 2 in a preprint submitted November 19, with Peiyun Hu listed as first author. The authors say the model estimates depth in meters for every pixel in a photo or video from a single camera, at the original input resolution. Peer review remains unconfirmed.

Apple published inference code with a command-line tool for photos and recorded video and Python examples that also cover live feeds. The repository omits training and evaluation code. Apple’s weight downloads include four model sizes, S, B, L and H+, and a shared focal-length estimator, with H+ set as the default in the command-line tool.

The Apple team trained one network on still images and video clips. The model uses one frame for a still and, for video, shares information across frames before producing a depth map for each one. A supplied or separately estimated camera focal length sets the scale in meters.

In tests on synthetic Sintel video, the authors report that Depth Pro 2-H+ scored 43.4 on boundary F1, a measure of how well predicted depth edges match reference edges, versus 36.3 for Depth Pro. They report scaled jitter of 1.39 for Depth Pro 2-H+ and 11.26 for Depth Pro. Lower jitter means the depth error changes less from frame to frame at the same visible points, with the error measured on a logarithmic scale.

The authors evaluated both models on the same 50-frame clips at each model’s own inference resolution. Depth Pro processed frames individually. The authors aligned predicted depth to the reference frame by frame for boundary F1 and once per video for jitter.

Apple also offers prebuilt Core ML packages with a fixed set of input sizes for Apple silicon Macs running macOS 15 or later. The inference code license permits use, modification and redistribution under its terms. The model license limits the weights and model derivatives to non-commercial scientific research and academic development and excludes product development.

More news