MetricAnything Revolutionizes Depth Learning Without Manual Calibration

A team of researchers has introduced MetricAnything, a new pretraining framework designed to improve metric depth learning. The system uses diverse 3D data without requiring manual calibration or specialized hardware. Its developers claim it sets a new benchmark in tasks like depth completion and monocular depth estimation.

The framework was created by Baorui Ma, Jiahui Yang of Li Auto Inc, Donglin Di, and their colleagues. It has already shown promise in autonomous driving and robotics, where precise spatial understanding is critical.

The team trained MetricAnything on roughly 20 million image-depth pairs sourced from over 10,000 different camera models. This large-scale approach follows the same scaling principles that have driven advances in modern foundation models. For the first time, the study demonstrates a clear trend: metric depth learning improves as the volume and diversity of training data increase.

One key innovation is the Sparse Metric Prompt technique. It works by randomly masking sections of depth maps, which helps separate spatial reasoning from biases tied to specific sensors or cameras. The result is a distilled, prompt-free model that outperforms existing methods in monocular depth estimation and camera intrinsics recovery.

Beyond standalone applications, the pretrained MetricAnything visual encoder can also enhance multimodal large language models. By integrating it, these models gain stronger spatial intelligence, making them more effective in tasks requiring 3D understanding.

To encourage further development, the researchers have open-sourced the framework. However, as of February 2026, no verified studies or practical tests outside autonomous driving and robotics have been documented. This limits comparisons in other potential use cases.

The release of MetricAnything provides a new tool for 3D perception research. Its ability to scale with data and improve depth estimation without manual calibration could accelerate progress in fields relying on spatial intelligence. The open-source nature of the framework allows other developers to explore its capabilities in real-world applications.