Data multiplication · physical AI

Capture once.
Train every condition.

Syntheta converts real factory, robotic, drone, and egocentric captures into labeled, validated training data. Generate controlled variation with Real2Robust, FunctaGen, and a clean passthrough baseline, while preserving source provenance and quality checks.

Original capture · clip-01-v2 · 10s 1072 × 480
Why this matters

Capture is the bottleneck.

5–30 min
of real capture to start a dataset run
Multi-class
industrial taxonomy with masks and boxes
Validated
variants with provenance and quality gates

Real-world capture is slow and rarely covers every operating condition. Syntheta takes the data you already have, restores and transforms it, then generates labeled variants for configurable industrial targets such as pallets, tools, robot arms, parts, defects, workers, and safety zones. NeuroDepth-T4 handles restoration, Real2Robust applies bounded domain variation, and FunctaGen provides learned synthesis.


Pipeline · seven stages

One capture.
A labeled dataset.

A source capture flows through restoration, target labeling, controlled generation, validation, dataset export, and model training. Each accepted sample keeps its source, recipe, seed, labels, and validation record.

FRAME01

Label once

A configurable target adapter labels the classes selected for the run, including industrial objects, people, robot components, parts, and safety regions.

detection
FRAME02

NeuroDepth-T4 restoration

EventCore V8.1 restores low-light frames using per-pixel brightness-change processing, with device, quality, and fallback state recorded in the manifest.

restoration
FRAME03

Branch A — R2R mutation

Real2Robust (R2R) applies domain-gap mutations — simulated weather, lighting shifts, sensor noise — to the restored frames. Labels carry forward from stage 1, no re-detection needed.

synthetic
FRAME04

Branch C — passthrough

The clean, un-mutated baseline — restored frames copied through with the same carried-forward labels.

baseline
FRAME05

FunctaGen — neural in-between frames

An implicit neural representation learns a compact scene-conditioned representation and produces source-anchored synthetic frames when its quality gates pass.

synthesis
FRAME06

Branch B — fresh detection + R2R

Generated frames are checked with target-aware detection and validation before they enter the training dataset. Fallback provenance is recorded instead of silently treating rejected output as ground truth.

synthetic
FRAME07

Summary

Every branch's output manifest rolls up into one summary file for the whole run.

manifest
BRANCHD

3D capture + human compositing — in progress

SfM reconstruction is real and working — 1,110 points reconstructed from 35 camera poses on a test run. Gaussian-splat training and SMPL-X human compositing into the rebuilt scene are next, currently blocked on GPU/CUDA and licensed SMPL-X/AMASS assets in this build environment, so it stops honestly rather than faking a render.

3D

Contact sheet

From one real capture

Inspect the transformation chain: real capture, restoration, mutation, generated labels, and 3D reconstruction.


Last validated run

Numbers, not adjectives

331
automated tests passing in the current build
6
controlled variation axes in generation recipes
6
suppression layers tuned in EventCore V8.1
30/255
brightness floor before dark-frame fallback
YOLO
detection, mask, and dataset export contracts
2D → 3D
source-anchored generation paths
6
pipeline stages from capture to evaluation
1
provenance contract across every modality
Built on

What's actually running

YOLOv8n-pose
NeuroDepth-T4 · EventCore V8.1
FunctaGen · INR + latent diffusion/flow
R2R mutation engine
YOLO / COCO dataset export
Gaussian splats · nerfstudio/gsplat (pending GPU)
SMPL-X + AMASS pose library (license-gated)

Get early access.

Bring a real capture and your target classes. We return an expanded dataset, labels, provenance, and a training baseline.

Sign up for beta