Learn · Synora

Train for the case the field rarely shows

The hardest cases are the rarest: a new pest, a disease strain, a new crop, low sun through dust. Synora generates photorealistic, labelled scenes of them so Folira and Targis are ready before the field shows them.

Synora · batch 0042, pest feeding damageExample
Real labelled images
180
Synthetic scenes
12,000
Held-out real test
90

ExampleBatch sizes are illustrative.

Capabilities

What Synora does

Generation

Labelled scenes at scale

Omniverse Replicator scenes with exact masks and boxes.

Randomisation

Light, dust and overlap

Sun angle, soil, residue, occlusion and motion blur are varied.

Cold start

New crops and regions

Starter data packs before a new region's first season.

Validation

Checked against reality

Synthetic-trained models are scored on real held-out images and Acreon outcomes.

World models

Field variation

Cosmos-based variation of real scenes (planned).

Licensing

For other ML teams

Datasets and pipelines for equipment makers (planned).

How it works

Synora in four steps

1

Find the gap

Acreon and Folira show which cases fail.

2

Build scenes

Plants, weeds and pests are placed in randomised fields.

3

Train

TAO fine-tunes the model on mixed real and synthetic data.

4

Prove

The new model must beat the old one on real images.

Specification

Planned specification

Planned values are design targets for the first release and will change with pilots.
ItemPlanned
Scene engineOmniverse Replicator
LabelsInstance masks, boxes, classes
ComputeDGX or cloud GPUs
Asset libraryCrops, weeds, pests, soils
ValidationReal held-out sets
OutputCOCO-style datasets

Status

Pre-launch

Nothing on this page is shipping today. Items marked planned are on the draft roadmap.

Draft roadmap

Questions

Questions about Synora

An agronomist or an engineer reads every message, not a sales queue.

Is synthetic data enough on its own?

No. It fills gaps. Every model is still tested on real images before it ships.

Can we license datasets?

Licensing to equipment makers and ML teams is planned after the first packs ship.

How do you avoid fake-looking scenes?

By measuring: synthetic data is kept only if it improves results on real held-out images.

Jump to a page or section