Depth Anything Small

Xenova/depth-anything-small-hf is a monocular depth estimation model that predicts a dense depth map from a single input image. It is the ONNX-exported version of the Depth Anything small model, designed for efficient inference in web and edge environments via Transformers.js. The model uses the DPT architecture with a DINOv2 backbone and contains 24.8 million parameters. It was trained on approximately 62 million images, including 1.5 million labeled and 62 million unlabeled images, as described in the paper "Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data" (arXiv:2401.10891), accepted at CVPR 2024. In zero-shot evaluations across six public datasets, the model demonstrates strong generalization to diverse scenes and image types. After fine-tuning with metric depth data from NYUv2 and KITTI, the Depth Anything family achieves state-of-the-art results.

Key capabilities

• Produces a per-pixel depth map from a single RGB image
• Supports zero-shot inference on unseen domains and randomly captured photos
• Optimized for client-side inference with ONNX weights

Output example

The model outputs both a raw depth tensor and a visualizable depth image. For example, given an input image of bread, the model generates a depth map as shown below:

The model is licensed under Apache-2.0 and is part of the Depth Anything release collection on Hugging Face, with over 70 community Spaces using the original PyTorch version.