Estimators

DepthFlow provides wrappers for SOTA Monocular Depth Estimation models with features like:

Cached results on disk to reduce import times and computational costs across runs.
Mitigate projection artifacts by fattening the edges, less foreground blending.

Usage¶

For using a specific #model, look for the depthflow.estimators package or command line help:

DepthScene¶

Python Command

# Already initialized as default
from depthflow.estimators.anything import DepthAnythingV2
from depthflow.scene import DepthScene

scene = DepthScene()
scene.estimator = DepthAnythingV2(model="small")

# See available commands
$ depthflow --help

# See estimator options
$ depthflow da2 --help

# Use DepthAnythingV2 small model
$ depthflow da2 --model small (...)

Standalone¶

Wrappers can also be used outside the scene, but always use and returns np.ndarray:

from depthflow.estimators.anything import DepthAnythingV2

estimator = DepthAnythingV2()
depthmap = estimator.estimate(image=...)

Models¶

-> Options below are roughly ordered by a combination of quality, size, and speed.

Depth Anything v2 ¶

Recommended and the default model

Showcase

Warning: Models other than Small are under CC BY-NC 4.0 (non-commercial + attribution) licenses.

Depth Anything v3 ¶

Currently awaiting a transformers safetensors release or alternatives

Likely available through the non-existing the *-hf version of depth-anything/DA3-SMALL
The pypi release contains way too many problematic and limiting dependencies, eg:
- open3d: Not used in Monocular Depth Estimation (MDE), has no Python 3.13+ prebuilt wheels.
- numpy<2: Version 1.26.4 was release in Feb 2025, the ecosystem already moved forward.
Fork at BrokenSource/Depth-Anything-3 reducing it, but I don't want to maintain a pypi package.
May add a runtime git+ install option, but that is frowned upon.