DINOv2 does depth: single-image estimation minus the bloat

What it does Depth Anything V2 is a foundation model for monocular depth estimation: feed it one RGB image and it returns a dense depth map. The project provides four scaled variants—Small through Giant—built on a DINOv2-DPT architecture. It handles both still images and video, with larger checkpoints yielding better temporal consistency for sequences.

The interesting bit The authors openly note that V1 accidentally used the last four DINOv2 layers for decoding; V2 switches to intermediate features purely to follow common practice, even though the change did not actually improve accuracy or details. That kind of honest architectural housekeeping is unusual in a field that usually treats every tweak as a major advance.

Key highlights

• Four model sizes (24.8M to 1.3B parameters), though the 1.3B Giant checkpoint is not yet released.
• The authors report faster inference, fewer parameters, and higher depth accuracy than SD-based depth models.
• Supports variable input resolutions; increasing the default 518px size yields finer-grained results.
• Broader ecosystem support than most research drops: Apple Core ML, Hugging Face Transformers, TensorRT, ONNX, ComfyUI, and even Android ports.
• The Small model is Apache-2.0 licensed; Base, Large, and Giant are CC-BY-NC-4.0.

Caveats

• The Giant model is listed as “Coming soon” with no release date provided.
• Only the smallest model (Small) is under a permissive open-source license; the larger checkpoints are restricted to non-commercial use.
• Predictions loaded via Hugging Face Transformers may differ slightly from the native implementation due to upsampling differences between Pillow and OpenCV.

Verdict Worth a look if you need off-the-shelf depth estimation for images or video without wrangling a full diffusion pipeline. Skip it if you require metric depth out of the box—relative depth is the default, and metric fine-tuning is handled separately.