Depth Anything 3

Depth Anything 3 is a state-of-the-art monocular depth estimation model designed to recover detailed 3D scene geometry from single images or video frames. Built on a plain transformer architecture and trained with a depth-ray representation, it delivers accurate, dense depth maps without requiring specialized 3D sensors or complex multi-view setups. The model supports diverse visual inputs, including indoor scenes, outdoor environments, synthetic data, and in-the-wild imagery, making it suitable for a wide range of computer vision and graphics applications.

Depth Anything 3 provides pre-trained checkpoints, inference scripts, and example notebooks, enabling developers and researchers to integrate high-quality depth estimation into their own pipelines with minimal effort. Typical use cases include 3D reconstruction, novel view synthesis, AR/VR content creation, robotics perception, autonomous navigation, and visual effects. The project emphasizes strong generalization, robustness to varying lighting and textures, and compatibility with common deep learning frameworks. It is particularly valuable for teams that need reliable depth information but lack extensive 3D capture infrastructure, allowing them to prototype and deploy geometry-aware AI systems efficiently. As an open research model, it also serves as a solid baseline for further experimentation and domain-specific fine-tuning.