Depth Pro 2 Delivers Sharp, Stable Video Depth

Depth Pro 2 depth map of a scene produced from a single camera image

Depth Pro 2 is a new computer vision model that estimates how far away every pixel in an image or video frame is, using only a single camera. Researchers posted the paper to arXiv on 19 November 2026 and say the model produces depth maps that are accurate in real-world units, crisp at object edges and steady from one video frame to the next.

Those three qualities have rarely come together in one system. The team says that is the gap Depth Pro 2 was built to close.

What monocular metric depth means

Monocular depth estimation predicts distance from one image, with no second camera or depth sensor. “Metric” means the output is in actual distances, such as meters, instead of a relative ordering where one object is simply nearer than another. Absolute distances matter for anything that has to act on what it sees, including robotics, augmented reality and video effects that place virtual objects behind real ones.

The model works on still images and on video, and it predicts depth at the input’s full native resolution, with no reduced-size pass that gets scaled up later.

The trade-off the authors set out to fix

After analyzing earlier monocular depth methods, the researchers found that most of them give up one quality to get another. Methods that look at several frames at once tend to be consistent over time, so depth doesn’t flicker as a clip plays. Methods that process a single frame at a time tend to draw sharper boundaries around objects.

Anyone who has watched a depth-based video effect knows why this matters. A sharp but jittery result shimmers on screen. A smooth but blurry one smears the edges of people and furniture into the background.

Depth Pro 2 tries to avoid the choice. The authors train one scalable network that handles both images and video, so a single model covers both cases instead of two separate pipelines.

What the researchers report

According to the paper, the model is metrically accurate, sharp and temporally consistent at once, and it outperforms prior work on all three measures. The team also releases a whole family of models. They say the family defines a new Pareto frontier across metric accuracy, sharpness, temporal consistency and runtime.

A Pareto frontier is the set of options where improving one measure means giving something up on another. Claiming a new frontier means that, at a given speed, no earlier model matched the new ones on all the quality measures together. Having several model sizes lets developers pick the point that suits their hardware and latency budget.

  • One network for both still images and video
  • Depth predicted at native resolution
  • A family of models that trade quality against runtime
  • Code and weights released with the paper

The results are the authors’ own, reported in a preprint that hasn’t been through peer review. Independent benchmarks will show how well the claims hold up on other data.

Why developers will care about the release

Publishing code and weights lets other teams test Depth Pro 2 on their own footage and fold it into existing tools. Depth models usually serve as a building block inside larger systems, so a drop-in replacement that is both sharper and steadier can improve a product without redesigning it.

For video work, the combination of per-frame sharpness and stable results across frames is the main selling point, because those are the two things earlier approaches kept trading against each other.

What to watch after the release

The first signs of how good Depth Pro 2 really is will come from outside tests. Researchers will likely compare it against other monocular depth models on standard benchmarks, and engineers will try it on messy real footage with fast motion, low light and reflective surfaces, where depth estimators often struggle.

Speed will be the other thing to check. The authors present runtime as one of the four axes on which their model family sets the frontier, so developers should see how the smaller models perform on modest hardware. If the results hold, Depth Pro 2 could become a common starting point for depth-aware video and robotics work.

Source: arXiv.org

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.