Depth Anything V2

A foundation model for monocular depth estimation (MDE): it delivers fine-grained, robust depth maps from a single image, operating 10x faster than SD-based models.

Depth Anything V2 is the next-generation foundation model for monocular depth estimation (MDE). It achieves superior, fine-grained depth predictions by leveraging a teacher-student distillation framework: a large teacher model is trained on 595K synthetic images, then generates pseudo-labels for over 62M unlabeled real-world images. This approach ensures greater robustness and detail than V1. Critically, it offers models from 25M to 1.3B parameters and boasts over 10x faster inference speed compared to models built on Stable Diffusion (SD), making it a highly efficient solution for real-time applications.

https://github.com/DepthAnything/Depth-Anything-V2