Key takeaways
- Apple researchers posted Depth Pro 2 to arXiv on November 19, 2026, with code and weights in four sizes from 24M to 872M parameters.
- The authors say one network now gives sharp and temporally steady metric depth for photos, video and live streams.
- The weights are licensed for non-commercial research only.
Apple researchers have released Depth Pro 2, a model that estimates how far away every pixel in a photo or video is, in meters, from a single ordinary camera. The paper, posted to arXiv on November 19, 2026, comes with code and weights in four sizes, from 24 million to 872 million parameters.
The problem it goes after is a trade-off. Depth models that look at one frame at a time draw sharp edges around hair, leaves and fingers, but their output flickers from frame to frame in video. Models that look at several frames stay steady but blur the edges. The eight authors, from Peiyun Hu to Vladlen Koltun, say one network now does both, and that the model family sets "a new Pareto frontier" across accuracy, sharpness, consistency and speed. Those are the team's own claims. The Apple research page gives no benchmark numbers, so the comparison has to be checked in the paper.
It matters to anyone building AR effects, video editing, robotics or 3D capture, because a depth map from a plain camera is cheaper than a depth sensor. The GitHub repository runs on Linux or an Apple silicon Mac with one command, works on photos, video files and live streams, and has Core ML versions for Apple hardware on Hugging Face. The catch is the license: the weights are for non-commercial research only, so a product can't ship them.
Depth Pro 2 follows Depth Pro, which Apple released in 2024 for single images. The new model adds video, which is where the flicker problem lived.
Sources
- Depth Pro 2: Sharp and Temporally Consistent Metric Monocular DeptharXiv
- Depth Pro 2Apple Machine Learning Research
- ml-depth-pro-2GitHub
- apple/DepthPro2Hugging Face
