Depth Anything V2

Depth Anything V2 refines the original monocular depth estimation model with markedly better fine-grained detail and robustness, while running faster with fewer parameters than diffusion-based depth methods. From a single image or video frame it predicts a relative depth map, and fine-tuned checkpoints deliver metric depth. Four encoder scales are offered: small at 24.8M parameters under Apache 2.0, plus base at 97.5M, large at 335.3M, and a planned 1.3B giant, with the larger variants licensed CC-BY-NC-4.0 for non-commercial use. Larger models also hold up better on video sequences, and variable input resolution can be used to extract extra detail. The model loads directly through Hugging Face Transformers and has been ported to Apple Core ML, TensorRT, and ONNX, with community integrations for ComfyUI and Android. Computer vision researchers and developers of robotics, 3D reconstruction, and image generation pipelines that need depth conditioning are the primary users.