Built on Synetic AI's synthetic rendering pipeline

The object detection model
trained on no real images

LYNX outperforms YOLOv8 on real-world data using 100% synthetic training. No AGPL. No licensing fees. Deployable from cloud to edge.

Copied to clipboard

View benchmarks

lynx-inference — orchard_row_14.jpg

Detections6

Latency12ms

Confidence0.94 avg

Train data100% synthetic

Real images seenZero

+24%higher mAP@50 than YOLOv8 trained on real data

0real images used during training

$0licensing fees for commercial deployment

Benchmarks

Faster. More accurate.
No licensing headaches.

Tested on 182 real-world orchard photos. Same dataset as the Blaga et al. 2025 whitepaper. LYNX was never shown a single real image during training.

Model	Train data	mAP@50	mAP@50:95	License
LYNXlynx-apple-v1	100% synthetic	0.697	0.344	Free / commercial
YOLOv12	Real world	0.628	0.322	AGPL · $75K/yr
YOLOv11	Real world	0.634	0.344	AGPL · $75K/yr
YOLOv8	Real world	0.561	0.243	AGPL · $75K/yr

Conf 0.1 · 182 real-world validation images · never seen during training · 3,280 synthetic training images

Read the whitepaper → Blaga et al. 2025

Quick start

Three lines
to deploy.

LYNX uses the same API surface as YOLO — drop-in replacement, no rewrites. Works with images, video streams, and RTSP feeds.

Install

Python 3.8+ required. Works in any virtualenv, Docker container, or conda environment.

Load a model

Pre-trained weights download automatically. Choose COCO-80, agriculture, or a custom class set.

Run inference

Pass a file path, URL, numpy array, or OpenCV frame. Returns bounding boxes, confidence, and class labels.

Python

CLI

Docker

# Install
pip install lynx

# Run inference
from lynx import LYNX

model = LYNX("lynx-coco.pt")
results = model("image.jpg")

# Results with bounding boxes
results.show()
results.save("output/")
print(results.pandas())

Models

Pre-trained models.
Ready to deploy.

LYNX-COCO

80 COCO object classes. General-purpose detection for products, people, vehicles, and common objects.

mAP@500.697

Classes80

Size31M params

Download Docs

LYNX-Agriculture

Apple, pear, crop disease, irrigation anomaly. Validated on orchard environments.

mAP@50TBD

ClassesCrop-specific

StatusTraining

Notify me

LYNX-Industrial

Surface defect detection, weld inspection, assembly QA. Trained for manufacturing environments.

mAP@50TBD

ClassesDefect types

StatusTraining

Notify me

Custom

Your objects. Your environment. Monthly weight updates as you send feedback data.

Pricing$1K/class/yr

BillingAnnual

UpdatesMonthly

Request model

Why LYNX

Built differently
from the ground up.

Better data

Trained on physics-accurate procedural renders from Unreal Engine with pixel-perfect annotations. No mislabeled images. No collection bias. No tail-risk edge cases. Pure signal at <$0.0001 per image.

No licensing risk

YOLO is AGPL-3.0. The moment you ship a product using it, you owe Ultralytics $75K/year. LYNX is commercially clean — use it in SaaS, embedded devices, or enterprise products today.

Deployable anywhere

Export to ONNX, TensorRT, or CoreML in one command. Native Jetson Orin support. ROS2 node included. Runs at 12ms on edge GPU, GPU-optional with ONNX runtime.

Custom models

Need a class
we don't have?

LYNX is trained on Synetic's synthetic data pipeline. Tell us what you need to detect — we generate the training data, train the model, and deliver monthly weight updates.

Synetic has 150+ peer-reviewed 3D assets and renders at scale on B200 infrastructure. Custom detection isn't a services engagement — it's a subscription.

$0forever

COCO 80-class model. Always free. No account required.

80 COCO object classes
Commercial use permitted
ONNX / TensorRT export
Community support

$1,000/ class / year

Custom classes trained on synthetic data generated specifically for your use case.

Any detectable object class
Monthly model weight updates
Physics-accurate training data
Dedicated Slack channel
Annual billing

Request a custom model

The object detection modeltrained on no real images

Faster. More accurate.No licensing headaches.

Three linesto deploy.