Apple Open-Sources Depth Pro 2: Metric Monocular Depth with Sharp Edges and Temporal Consistency
Apple has released and open-sourced Depth Pro 2, a scalable monocular depth estimation network that estimates metric (real-scale) depth per pixel at native resolution from a single photo, an entire video, or live footage. It is available on Hugging Face with an online demo. The model comes in four sizes: S (24M parameters / 97MB), B (97M / 386MB), and more, with the largest at 872M parameters.
Confirmed
- According to Apple's official account, existing monocular depth estimation methods face a trade-off: multi-frame approaches favor temporal consistency while single-frame approaches favor sharp edges, making it hard to achieve both; Depth Pro 2 delivers metric depth accuracy, sharp object edges, and temporal consistency across video frames with a single network
- The model supports native-resolution input, and @akhaliq noted its depth maps show sharp object edges and stable timing across frames
- Beyond the main model, Apple also released DepthPro2-focal, a focal length estimator, on Hugging Face: a compact network with only 23.5M parameters that predicts camera focal length from a single photo or video frame; its pipeline is depth-estimation, with tags indicating support for focal length estimation, field-of-view, and other capabilities, useful for vision tasks requiring focal length information such as depth estimation and 3D reconstruction
Why it matters
- Metric depth (directly outputting real-world scale) combined with temporal consistency makes the model suitable not only for single-image 3D understanding but also for video and real-time scenarios (e.g., AR, robotics, 3D reconstruction)
- Sizes ranging from 24M to 872M parameters cover needs from lightweight on-device deployment to high-accuracy applications
- The bundled compact focal length estimator lowers the barrier to obtaining camera intrinsics for tasks like single-image 3D reconstruction
2026-11-20 ~ 2026-11-20 · 6 related posts
Primary sources
- [source] Apple ships Depth Pro 2 on Hugging Face with focal length and FOV estimation — apple · 2026-11-20
- [source] Apple releases Depth Pro 2: sharp, temporally consistent metric monocular depth, up to 872M params — _akhaliq · 2026-11-20
- Apple releases compact 23.5M-parameter focal length estimator for Depth Pro 2 on Hugging Face — bdsqlsz · 2026-11-20
- [source] Apple's Depth Pro 2 delivers sharp, temporally consistent metric monocular depth — apple · 2026-11-20