golynx

Object detection where every measurement arrives with its own provenance — so you can always tell a measured metre from an assumed one.

pip install golynx

import lynx

with lynx.open_standard("yolo26n.onnx") as model:
    frame = model.predict("street.jpg")
    for d in frame.detections:
        print(d.class_name, d.box.trust.confidence)

The distribution is golynx; the module is lynx.

Two ways in

A model you already have — .lnx, .lnxp, or a raw ONNX export. No registry, no network. Dispatches on the file's contents, not its extension:

model = lynx.open_standard("yolo26n.onnx")

A catalogue model by slug — fetched, licence-checked, cached:

model = lynx.open("your-model-slug")

Metadata a foreign graph doesn't carry is deduced when it opens. probe_config() works out what an unfamiliar export needs by running it. autotune() returns ranked evidence without applying anything, so a coin-flip stays visibly a coin-flip.

Every value carries a Trust

The same fields on a box, a depth, a mask and a keypoint:

source separates what the model was trained to report (TRAINED, TRAINED_RELATIVE, TRAINED_SIZE, TRAINED_RANGE) from what the SDK worked out for you (DERIVED_DEPTH, DERIVED_DECLARED_HEIGHT, DERIVED_ASSUMED_FOCAL, DERIVED_POPULATION, DERIVED_ASSUMED_POSTURE).

A bare detection export has no depth head, yet every detection still reports a depth — because the SDK knows roughly how big a bus is. That reading is honest about itself: it comes back DERIVED_ASSUMED_FOCAL with a wide margin. Hand it camera intrinsics and the same reading becomes DERIVED_DECLARED_HEIGHT with a much tighter one.

Heads

BOUNDING_BOX, ORIENTED_BOUNDING_BOX, SEGMENTATION, INSTANCE_SEGMENTATION, DEPTH, POSE, CLASSIFICATION, REID, TEXT_RECOGNITION, AMODAL_BOX, YAW_3D, METRIC_SIZE, RANGE, DENSITY. A model exposes what it declares; ask for the tasks you want.

Geometry

cam = lynx.Intrinsics(fx=800, fy=800, cx=405, cy=540)

for d in frame.detections:
    p = d.position_3d(cam)          # metres, +z forward
    print(d.class_name, p.z, d.depth.value, d.depth.trust.source)

Depth, 3D position, real-world size and surface-to-surface distance are computed in the SDK rather than left to the caller.

Platforms

Python 3.10 – 3.14 on Linux x86-64 (manylinux_2_34), macOS Apple silicon, and Windows x64.

CUDA and Jetson JetPack 6 builds are too large for PyPI and are served separately:

pip install --extra-index-url https://syneticai.github.io/lynx/simple golynx-gpu

© Synetic AI