Image forensics at volume: up to 100,000 images an hour.
Every photo your programme receives is checked for authenticity, quality and reuse before it costs you anything: who took it, on what, whether a machine made it, and whether you have seen this person, or this exact image, before. Automated, auditable, and fast enough to keep pace with intake.
Measured, not estimated
What this line actually covers
Large programmes run on images submitted by people: partner onboarding, KYC and account sign-up, insurance and warranty claims, marketplace listings, field and retail audits, user-generated campaigns. At volume, a meaningful share of those images is reused, edited, AI-generated, or simply unusable, and nobody can look at a hundred thousand photos by eye.
FRAM3’s forensics pipeline examines every image and reports what it finds, per image and per account.
What every image is checked for
Provenance and metadata
Capture device, capture time and GPS location where present, plus whether the file has been through an editor.
AI-generated and AI-edited images
Content-provenance markers (C2PA and IPTC digital-source-type) that identify fully synthetic images and images altered with generative tools, including “clean-up” style edits on phones. Validated on both real-world submissions and AI-generated statics.
Duplicates and reuse, in three tiers
- Identical files: the same image submitted more than once.
- Near-duplicates: the same picture re-saved, resized, cropped or re-compressed.
- Same person: different photos of the same face submitted under different accounts or IDs, found with face-identity matching and grouped so each cluster can be reviewed as one case.
Quality and fitness for use
Resolution, blur, lighting and backlight, number of faces, framing, off-brand or competitor content, and broken or mislabelled files (a video uploaded as a photo, a truncated file).
Recovery, not just rejection
Where a photo fails on quality alone, identity-preserving enhancement can often bring it back into use. Every recovered image is checked against the original so the person stays the same person.
Built for the size of the problem
Throughput. Up to 100,000 images an hour, scaled on GPU infrastructure to the size of the archive or the daily intake.
Two modes. A one-time audit of an existing archive, or continuous screening at intake so a bad image is caught before anything is built on it.
Evidence, not scores. Each finding comes with its reason: the matching file, the cluster of accounts that share a face, the provenance marker. Your team can verify it.
Your data stays yours. Images are read where they live; anomalies are reported back into your own storage, organised for review.
What one audit found
For a national partner programme, a single run over 127,000 submitted photos took under 30 minutes and surfaced:
- tens of thousands of photos reused across different accounts, many byte-for-byte identical;
- over a thousand images carrying conclusive AI-generation or AI-editing provenance;
- the same person appearing under multiple IDs, grouped into reviewable clusters.
Most of these had already passed conventional quality checks.
Who this is for
Any organisation that makes decisions or spends money on the strength of a submitted image: financial services and insurers (onboarding, KYC, claims), consumer brands running dealer, distributor or loyalty programmes, marketplaces and classifieds, telecom and retail field operations, and marketing teams running large user-generated or personalised campaigns.
Where FRAM3 sits
Single-image checking tools answer “is this one photo real?”. Volume programmes need a different question answered: “across everything we have received, what is reused, what is synthetic, and who appears more than once?” That is a cross-corpus problem, and it is the one our pipeline is built for. It is proven inside our own production line, where every photo is screened before a video is made from it.
What we deliver
- Forensic analysis of up to 100,000 images an hour
- AI-generated and AI-edited image detection from provenance markers
- Three-tier duplicate detection, up to the same face across different accounts
- Capture metadata extraction: device, time, location
- Automated quality gates, with identity-preserving recovery for usable images
- Archive audits and continuous screening at intake
- Findings delivered as reviewable cases, traceable to the source file
One engine, sharpened from every direction.
Sitting on a mountain of submitted images?
Tell us the volume and what you need to know. We will tell you what is in there.