Depth Anything V2 is a foundation model for monocular depth estimation that significantly outperforms its predecessor in fine-grained detail and robustness. Published at NeurIPS 2024 with 7.8k+ GitHub stars and Apache 2.0 licensing, it has become the go-to solution for generating high-quality depth maps from single RGB images, with models ranging from 25M to 1.3B parameters. ## The Depth Estimation Problem Monocular depth estimation — predicting the distance of every pixel in a scene from a single 2D image — is one of the core challenges in computer vision. Unlike stereo systems that use two cameras, monocular approaches must infer 3D structure from 2D cues alone. This capability is essential for autonomous driving, robotics, augmented reality, 3D scene reconstruction, and video editing. Depth Anything V2 tackles this by training on a massive combination of labeled and unlabeled data, creating a versatile foundation model that generalizes across diverse scenes without domain-specific fine-tuning. ## Architecture and Training ### DINOv2 Backbone Depth Anything V2 builds on the DINOv2 vision transformer as its encoder backbone, inheriting powerful visual representations learned through self-supervised training. The decoder architecture has been refined for more precise depth prediction, particularly at object boundaries and fine-grained structural details. ### Synthetic-to-Real Training Strategy A key innovation in V2 is the improved training strategy. The model first trains on large-scale synthetic data with precise ground-truth depth labels, then bridges the synthetic-to-real domain gap through a carefully designed knowledge distillation process. This approach avoids the noise inherent in real-world pseudo-labels while maintaining strong generalization to natural images. ### Model Variants The model family spans a wide range of computational requirements: the smallest variant (25M parameters) runs efficiently on edge devices, while the largest (1.3B parameters, based on DINOv2-G) delivers the highest accuracy for applications where compute is not a constraint. This flexibility makes Depth Anything V2 suitable for everything from mobile AR applications to cloud-based 3D reconstruction pipelines. ## Key Capabilities ### Video Depth Anything Released in January 2025, Video Depth Anything extends the model to generate temporally consistent depth maps for video sequences, handling clips over 5 minutes long. This addresses the flickering and temporal inconsistency that plagued frame-by-frame depth estimation approaches. ### Prompt Depth Anything The Prompt Depth Anything extension supports 4K resolution metric depth estimation when low-resolution LiDAR data is available as a prompt. This hybrid approach combines the global understanding of the neural network with sparse but accurate sensor measurements, achieving metric-scale accuracy suitable for robotics and autonomous driving. ### Broad Integration Depth Anything V2 is integrated into Hugging Face Transformers, Apple Core ML Models, ONNX Runtime, and TensorRT. The Transformers integration means developers can load and run the model in just a few lines of Python code, while the Core ML export enables efficient on-device inference on Apple hardware. ## Benchmarks and Performance Depth Anything V2 achieves state-of-the-art results across standard benchmarks including NYU Depth V2, KITTI, and ETH3D. The model demonstrates significantly sharper depth boundaries, better handling of thin structures, and more robust performance in challenging conditions like reflective surfaces, transparent objects, and low-light scenes compared to V1. Compared to Stable Diffusion-based depth models like Marigold, Depth Anything V2 is faster in inference, has fewer parameters, and achieves higher depth accuracy on standard benchmarks. ## Practical Applications The model serves diverse use cases across industries. In robotics, it provides real-time depth perception for navigation and manipulation tasks. In AR/VR, it enables realistic object placement and occlusion handling. In 3D reconstruction, it generates dense depth maps for photogrammetry pipelines. In video production, it powers depth-aware effects like selective focus blur, relighting, and 3D parallax. ## Limitations As a monocular approach, absolute depth scale cannot be determined without additional calibration or reference measurements. Performance degrades on highly unusual viewpoints or scene types that are underrepresented in training data. Very small objects or extremely fine textures can still challenge the model. Real-time performance on edge devices requires using the smallest model variants, which sacrifice some accuracy. ## Community and Development The project maintains active development with extensions like Video Depth Anything and Prompt Depth Anything demonstrating continued innovation. With nearly 800 forks and adoption across major ML frameworks, Depth Anything V2 has established itself as the standard foundation model for monocular depth estimation in both research and production settings.