Depth Anything V2 is a monocular depth estimation model that predicts depth maps from single images.
The tool addresses the problem of inferring three-dimensional scene structure from monocular input by training a foundation model capable of fine-grained depth prediction across diverse scenes. It operates as a neural network that processes images and outputs corresponding depth maps, with the approach emphasizing robustness and accuracy in both relative and metric depth estimation modes.
Developers should choose this tool if they need depth estimation without stereo input or structured light sensors. It suits applications ranging from 3D reconstruction and autonomous navigation to augmented reality and robotics. The project offers four model variants scaling from 24.8M to 1.3B parameters, allowing trade-offs between inference speed and accuracy depending on deployment constraints. Compared to diffusion-based depth models, the tool provides faster inference, fewer parameters, and higher depth accuracy. The project has expanded beyond single-image estimation to support video depth prediction for extended sequences and metric depth refinement when low-resolution LiDAR prompts are available.
The project maintains active development with recent releases of complementary tools including Video Depth Anything for temporal consistency across long video sequences and Prompt Depth Anything for metric depth at 4K resolution. Integration into major frameworks including Hugging Face Transformers and Apple Core ML Models indicates broad ecosystem adoption. The codebase includes a dedicated benchmark dataset and provides pre-trained checkpoints across all model scales, with smaller metric depth variants available alongside the base relative depth models.
