
Depth Pro 2: Sharp, Metric, Temporally Consistent Depth from a Single Network

Depth Pro 2 is a method for dense metric monocular depth estimation that handles both still images and video at full resolution. The authors identify a common trade-off in prior work: multi-frame methods tend to favor temporal consistency, while single-frame methods favor boundary sharpness. Instead of accepting that split, they train one scalable network that accepts single photos, whole videos, or a live camera feed and produces depth for every pixel in meters at the input’s native resolution. The resulting depth is reported to be metrically accurate, sharp at object edges, and steady from frame to frame simultaneously, outperforming earlier work on all of those dimensions.
The project page says the released family of models defines a new Pareto frontier in metric accuracy, sharpness, temporal consistency, and runtime. Code and model weights are released. Related readings link to Depth Pro, the earlier zero-shot metric monocular depth foundation model, and SHARP, a separate system for photorealistic single-image view synthesis through regression of a 3D Gaussian representation.

