Depth Anything V1

Depth Anything is a foundation model for monocular depth estimation, presented at CVPR 2024 by researchers from HKU, TikTok, CUHK, and ZJU. Its training recipe combines 1.5 million labeled images with more than 62 million unlabeled ones, which gives it strong generalization to unseen scenes from a single photograph. Three variants are released, small at 24.8M parameters, base at 97.5M, and large at 335.3M, producing relative depth maps, and fine-tuned checkpoints provide metric depth for NYUv2 and KITTI. The model outperforms MiDaS v3.1 across benchmarks including Sintel, ETH3D, and DIODE, and its encoder transfers well to tasks like semantic segmentation. Adoption in the generation ecosystem is broad: a depth-conditioned ControlNet is provided, and integrations exist for Hugging Face Transformers, ONNX, TensorRT, ComfyUI, InstantID, and InvokeAI, plus community ports for Android and Jetson. The code is released under Apache 2.0, and V2 has since superseded it.