# Depth Pro 2 — Apple's open depth model for sharp, steady video depth

> One Apple network gives depth maps that are both sharp at object edges and steady from frame to frame.

- Org: Apple
- Date: 2026-11-19
- Category: model, paper
- Source: https://huggingface.co/apple/DepthPro2
- Canonical page: https://ai-tldr.dev/releases/apple-depth-pro-2/

## In short

Apple's Depth Pro 2 estimates metric depth in meters for photos, video and live camera feeds at full resolution. It ships in four sizes, from 24M to 872M parameters, with code and research-only weights.

## What it is

Depth Pro 2 is the successor to Apple's Depth Pro: a family of models that predicts how far away every pixel is, in meters, from an ordinary photo, a video or a live feed. It comes in S (24M), B (97M), L (334M) and H+ (872M) sizes, plus a shared 23.5M focal-length estimator for when the camera's focal length is unknown, and sits in the Hugging Face trending top 10.

## How it works

A multi-view transformer reads one or more frames together. After the first third of its blocks it alternates attention inside each frame with attention across frames, and a DPT decoder then predicts an inverse depth map for every frame at native resolution.

## Why it matters

The paper argues that earlier depth models forced a choice: multi-frame methods stay consistent over time, single-frame methods keep edges sharp. Apple reports that Depth Pro 2 beats both kinds on boundary sharpness and jitter across the Sintel and Spring benchmarks. The weights are licensed for non-commercial research only.

## Who it's for

computer vision researchers, 3D and video effects developers

## Try it

uv run depth-pro-2 predict -i assets/example.mp4 -o out/

## Links

- [Model Card](https://huggingface.co/apple/DepthPro2)
- [arXiv Paper](https://arxiv.org/abs/2611.26603)
- [Code](https://github.com/apple-aiml-research/ml-depth-pro-2)
- [Project Page](https://apple-aiml-research.github.io/ml-depth-pro-2/)
- [HF Trending](https://huggingface.co/models?sort=trending)
