Depth Anything V2

Depth Anything V2, released June 14, 2024 by researchers at HKU and TikTok and presented at NeurIPS 2024, is a foundation model for monocular depth estimation: predicting per-pixel depth from a single image. It improves on V1 with three changes: labeled real images are replaced entirely with synthetic images where perfect depth is available, a much larger teacher model is used, and student models are trained via large-scale pseudo-labeled real images. The authors report more than 10x faster inference and higher accuracy than Stable-Diffusion-based depth models, with variants from 25M to 1.3B parameters to cover phones to servers.

Code, models, demos and a new precise-annotation evaluation benchmark were all released openly, and V2 quickly became a default depth backbone in tools like ControlNet pipelines and ComfyUI, with Transformers integration following in July 2024. A January 2025 follow-up, Video Depth Anything, extended it to consistent depth over long videos.