Apple's New Model Measures the World in Metres From a Single Photo
Apple has released a model that looks at an ordinary photo or video and tells you how far away everything in it is — in real metres. Depth Pro 2 was announced on November 19, 2026, in the paper "Depth Pro 2: Sharp and Te
Apple has released a model that looks at an ordinary photo or video and tells you how far away everything in it is — in real metres. Depth Pro 2 was announced on November 19, 2026, in the paper "Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth" (arXiv 2611.26603). It estimates metric depth for every pixel, from a single photo, a whole video, or a live camera feed, at the input's full resolution.
A flat image does not record distance; a model has to infer it from clues like size, perspective, occlusion, and texture. Depth Pro 2 does that with one scalable network — a multi-view transformer paired with a DPT decoder — that handles both still images and video, reporting depth in metres rather than in relative terms.
One network for photos and video
Using one network for both images and video keeps things simple, and it helps video most of all: many depth models treat each frame on its own, which can make the depth flicker as the scene moves, while a network that looks across views keeps its answers steady from frame to frame.
Pick your size
Depth Pro 2 comes in four sizes, each with its own repo on Hugging Face:
- S: 24 million parameters, 97MB
- B: 97 million parameters, 386MB
- L: 334 million parameters, 1.3GB
- H+: 872 million parameters, 3.5GB — the default, and the best quality
The small models run where memory and compute are tight; the large ones trade size for quality. A shared focal-length estimator (23.5 million parameters, 94MB) predicts the camera's focal length when it is unknown, which matters because it shapes how the model reads perspective.
Video has two modes
Video can run in two ways. The first uses buffered clips of 16 frames with a 4-frame overlap, so each clip shares frames with the next and the depth stays smooth across the join. The second is per-frame streaming, where frames are handled one at a time as they arrive. Buffered mode suits recorded video; streaming suits a live feed, where the next frame has not arrived yet and the model has to keep up.
It runs on your Mac too
Depth Pro 2 runs on an NVIDIA GPU with CUDA in bfloat16, on an Apple silicon GPU through MPS, or on a CPU in float32. Core ML exports are on Hugging Face as well, and the repo includes a Mac live-depth demo — point a camera at a scene and watch the depth map appear in real time.
Apple says the model beats prior work on two measures: sharpness, judged by Boundary F1, and temporal consistency, judged by Jitter. Sharpness is about edges — a good depth map keeps a crisp line where one object ends and another begins. Temporal consistency is about stability — the depth of a still object should not wobble as the video plays.
The licence is the catch
Here is the part to read before building anything: Depth Pro 2 is released under the Apple Machine Learning Research Model licence, which allows research use only. It is not a commercial licence, so the model cannot be used in a product that makes money.
That still leaves plenty of room for researchers, who can use it freely, compare it with their own work, and build on what it shows. The code is on GitHub in the apple-aiml-research/ml-depth-pro-2 repo, the weights are on Hugging Face at apple/DepthPro2, and Apple's research page is at machinelearning.apple.com.
For researchers, the next step is simple: try the demo and test it on your own footage. For everyone else, Depth Pro 2 is a clear sign of where single-camera 3D is heading — and a reminder to read the licence before building on it.
Read the full article with all the details on Rohan Tech Official: https://rohantechofficial.com/tech-news/apple-depth-pro-2-metric-depth/
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.