Synthetic Data for Redaction Detectors

Problem

Redaction detectors need training imagery for exactly the situations that are hardest to collect: rare objects, awkward angles, edge cases that appear a handful of times across an entire corpus. Waiting to capture and label enough real examples was the binding constraint on how fast a detector could reach production — roughly eight months per cycle.

Approach

A synthetic-data pipeline built on diffusion-based image synthesis, paired with VLM-guided auto-labeling so generated imagery arrived already annotated rather than queuing for human labelling. Generation targeted the specific failure modes the detectors were losing on, instead of producing volume for its own sake.

Result

Research-to-production dropped from 8 months to 2–3. The pipeline came to supply 30% of all training data for the redaction detectors, and the models it fed improved precision and recall by 5–10%, going on to serve 40% of customer redaction usage.