Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth
Depth Pro 2 estimates metric depth for every pixel from an ordinary camera: a single photo, a whole video, or a live feed, at the input's full resolution. Its depth is sharp at object edges and steady from frame to frame.
Models
Each size is its own repository with its weights. Larger sizes give sharper, steadier depth; smaller ones run faster.
| Model | Repository | Parameters | Download size |
|---|---|---|---|
| S | apple/DepthPro2-S | 24M | 97 MB |
| B | apple/DepthPro2-B | 97M | 386 MB |
| L | apple/DepthPro2-L | 334M | 1.3 GB |
| H+ | apple/DepthPro2-H | 872M | 3.5 GB |
All sizes share a focal length estimator, apple/DepthPro2-focal (23.5M parameters, 94 MB), which loads with each size and predicts the focal length when it is unknown.
Quick start
pip install "git+https://github.com/apple-aiml-research/ml-depth-pro-2"
from functools import partial
import torch
from depth_pro_2 import DepthPro2, load_rgb, upright
from depth_pro_2.model import to_normalized_tensor
# apple/DepthPro2-S, -B, -L or -H; bfloat16 is how the models run on a GPU.
model = DepthPro2.from_pretrained("apple/DepthPro2-H", precision=torch.bfloat16)
model = model.to("cuda")
transform = partial(to_normalized_tensor, device=torch.device("cuda"))
Image
# f_px is the focal length recorded in the photo, or None.
image, f_px = load_rgb("photo.jpg")
depth = model.infer(transform(image), f_px=f_px)["depth"] # meters, (H, W)
Video
import av
# 16 frames at a time: one prediction per frame, in order.
video = av.open("video.mp4")
frames = (transform(upright(f)) for f in video.decode(video=0))
for prediction in model.infer_video(frames):
depth = prediction["depth"] # meters, (H, W)
Depth is infinite where the scene is infinitely far, such as the sky. The code has the command line and a live-feed example.
Results
Sharpness and temporal consistency
Depth Pro 2 outperforms prior work operating on a single frame or multiple frames on sharpness (Boundary F1 score) and temporal consistency (Jitter). Higher F1 and lower jitter are better, so the top right is best. Results averaged over Sintel and Spring datasets.
Photos
Top row: the input photos. Bottom row: Depth Pro 2's depth, red near and blue far.
Method
A multi-view transformer processes one or more frames jointly, and a DPT decoder predicts an inverse depth map for each frame. β After the first third of its blocks, the transformer alternates attention within each frame and attention across frames. β‘ The decoder assembles tokens from several blocks into feature maps at 1/32 to 1/4 of the input resolution.
License
The weights are released under the Apple Machine Learning Research Model license, for research use only: non-commercial scientific research and academic development.
Citation
@article{DepthPro2:2026:arxiv,
title = {Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth},
author = {Peiyun Hu and Marko Mihajlovic and Lyndon Duong and Lahav Lipson and Wei Dong and Ama\"{e}l Delaunoy and Stephan R. Richter and Vladlen Koltun},
journal = {arXiv preprint arXiv:2611.26603},
year = {2026},
url = {https://arxiv.org/abs/2611.26603},
}


