Depth Anything V2 Small

Depth-Anything-V2-Small-hf is a monocular depth estimation model that produces fine-grained and robust depth predictions from a single image. It is built on the DPT architecture with a DINOv2 backbone and trained on 595K synthetic labeled images and over 62 million real unlabeled images.

Key Strengths

• Produces more fine-grained details than Depth Anything V1.
• More robust than Depth Anything V1 and Stable Diffusion-based models such as Marigold and Geowizard.
• Over 10x faster and more lightweight than SD-based alternatives.
• Highly effective when fine-tuned for metric depth estimation.

Performance and Comparison

As shown in the Depth Anything V2 paper, the model achieves all six preferable properties—fine detail, transparent objects, reflections, complex scenes, efficiency, and transferability—whereas Marigold and Depth Anything V1 each lack multiple properties. The authors also constructed the DA-2K evaluation benchmark with precise annotations and diverse scenes to address limitations of existing test sets.

Scalability and Publication

Depth Anything V2 offers model scales from 25M to 1.3B parameters. The work was accepted at NeurIPS 2024. For the full methodology, see the paper: Depth Anything V2.