Depth Anything

Depth Anything V2 is the state-of-the-art monocular depth estimation (MDE) foundation model, leveraging a massive 62M+ real unlabeled image dataset for superior zero-shot generalization.

This model establishes a new benchmark for Monocular Depth Estimation: it accurately predicts depth from a single 2D image. Depth Anything V2 achieves this through a robust teacher-student framework, training on a massive dataset of 595K synthetic labeled images and over 62 million pseudo-labeled real-world images . Its transformer-based architecture (using DINOv2 encoders and a DPT decoder) ensures fine-grained detail and exceptional robustness across diverse scenes . The result is a highly efficient model, offering up to 10x faster inference than Stable Diffusion-based competitors, with model scales ranging from 25M to 1.3B parameters to suit any scenario .

https://depth-anything.github.io/v2/