Depth Anything V1

Depth Anything V1 is the original January 2024 monocular depth estimation foundation model from Lihe Yang and collaborators (HKU and TikTok), published as arXiv 2401.10891 and presented at CVPR 2024. Rather than inventing new modules, it scaled a data engine to collect and automatically annotate 62M+ unlabeled images on top of 1.5M labeled ones, using aggressive augmentation and auxiliary semantic supervision to learn robust representations. It set new zero-shot standards across six public datasets and improved depth-conditioned ControlNet as a side effect.

V1 is the predecessor of the more capable Depth Anything V2 (June 2024), but it remains the clean, Apache-2.0-licensed baseline the community built on — with checkpoints in small/base/large sizes, a Hugging Face demo, and wide integration into 3D, video, and robotics pipelines. The honest picture: it predicts relative (not metric) depth out of the box — you need the separate fine-tuned metric variants for NYUv2/KITTI-style absolute depth — and like all monocular estimators it can struggle with transparent surfaces, mirrors, and thin structures.