Depth Anything v2

Depth Anything v2 is one of the most advanced models for extracting depth maps from a single image (monocular depth estimation). Developed as a successor to the original Depth Anything model, the v2 version offers significant improvements particularly in fine details and edge accuracy. Created by researchers at the University of Hong Kong and TikTok's research team, this model has established a new performance standard in the depth estimation field, significantly improving the reliability of monocular depth estimation in practical real-world applications.

The model is trained with a smart combination of synthetic and real-world data using a carefully designed curriculum. This approach, leveraging labeled synthetic data and unlabeled real-world data, enables the model to both produce accurate depth values and generalize to diverse real-world scenes without domain-specific fine-tuning. It delivers consistent results across all types of scenes including indoor, outdoor, natural, and urban environments. The v2 training strategy leverages high-quality depth labels from synthetic data while strengthening the model's generalization capacity through the diversity of real-world data. This hybrid training approach improves both absolute depth accuracy and relative depth ordering simultaneously.

Depth Anything v2's most notable improvements over the previous version are seen in edge precision and fine detail preservation throughout the depth map. Object boundaries, thin structures, and complex geometries are represented with sharper and more accurate depth transitions that closely match ground truth. This improvement makes a significant difference particularly in applications where edge quality is critical, such as 3D scene reconstruction and augmented reality overlay alignment. The model also demonstrates improved performance in traditionally challenging scenarios such as reflective surfaces, transparent objects, and repetitive textures. It produces consistent predictions even in regions with ambiguous depth cues such as skies, water surfaces, and glass.

Available in three model sizes—Small, Base, and Large—each optimized for different deployment scenarios. The smallest variant is suitable for real-time applications with minimal latency requirements, while the Large variant provides the highest accuracy for offline processing tasks. Its DINOv2-based image encoder forms the foundation of the model's strong feature extraction capacity, utilizing rich visual representations from self-supervised pretraining. Each size is optimized for different computational budgets and latency requirements, allowing developers to select the most appropriate balance for their application needs and hardware constraints.

Depth Anything v2's practical applications are extremely broad and span multiple industries. 3D scene creation, augmented reality effects, robotic navigation, autonomous vehicle perception, and photo editing are among the most common applications driving adoption. In photography apps, it provides essential input for portrait mode background blurring, depth-based focus effects, and layered editing capabilities that simulate professional camera results. It can also be used as input for 3D reconstruction techniques like NeRF and Gaussian Splatting, significantly improving the quality of these reconstruction methods. In video games and film production, it is used for virtual camera effects and depth-based post-production compositing.

Released as open source, the model is accessible via Hugging Face and available in PyTorch, ONNX, and TensorRT formats for flexible deployment. It can be quickly tested through a Gradio-based demo interface for evaluation purposes. It offers frame-by-frame processing support for video depth estimation and can be combined with temporal consistency techniques to achieve smooth video depth maps suitable for production use. The model has rapidly gained acceptance as a standard depth estimation solution in the computer vision research community, serving as a foundation for numerous downstream applications and research directions.