Video depth has long involved a trade-off: processing frames together improves temporal stability, while single-frame methods tend to preserve sharper edges. The key idea here is to train one scalable model for both settings, so depth can remain crisp at object boundaries without flickering from frame to frame.
Key Capabilities
- It predicts depth in meters at the input’s native resolution, making fine boundaries more useful for downstream vision tasks.
- A multi-view transformer can process one frame or several jointly; cross-frame attention helps steady video predictions, while training on single images preserves the ability to infer metric scale from a photo.
- A shared focal-length estimator handles images without known camera intrinsics. When focal length is available, it can instead be supplied to recover metric depth.
- Four sizes, from 24M to 872M parameters, offer a practical speed-versus-quality choice. The reported streaming results range from 53 fps for S to 8 fps for H+ at 1080p on a B200.
Who It’s For
A good fit for research and vision pipelines that need dense, metric depth from ordinary cameras, especially when video stability and edge detail both matter. Look elsewhere if you need commercially licensed weights, or if your task requires geometry beyond monocular depth estimation; the model weights are limited to non-commercial scientific research and academic development.