Omega Lens

Seeing is no longer believing. Omega Lens is a sovereign recognition engine that judges whether an image, a voice, or a video was generated or manipulated, and can restore the missing context around a clip that was cut to mislead. It runs on your own hardware, using open-weight models, so no media ever ships to a third-party detector.

Why we built it

Generated media is now cheap enough to weaponize at scale: cloned voices for fraud calls, face-swapped video for extortion and disinformation, synthetic images to prop up a clone site. The tools that spot these are usually cloud services, which means handing your most sensitive media to a third party. Omega Lens was built so the detection can run where the compute already lives, on the operator's own box, and so the same trained-model brain can serve every Omega Point product without any of them shipping evidence off-site.

The problem it confronts

Two problems, really. First, trained neural detectors do not belong in a lightweight edge worker, they need real model runtime and weights, so the heavy judgment has to live somewhere with compute. Second, a clip can lie by omission: a real ten-second cut, pulled from a longer source and stripped of what came before and after, can mislead as badly as an outright fake. Omega Lens addresses both, the trained-model verdict and the missing-context recovery.

How it works

• Image, audio, and video recognition. Fetches the media under an SSRF guard, then routes it to the right detector: AI-image and face-manipulation classifiers, voice-spoof and clone classifiers, and frame-sampled aggregation for video. Devices are auto-detected, models are lazy-loaded.
• Context restoration. For a suspicious clip it runs edit forensics (hard cuts, audio splices, very short duration), a verbatim timestamped transcript, transparent rule-based technique and arousal flags, and source recovery by perceptual fingerprint against a sovereign corpus you index, with clearly-flagged external reverse-search only as a fallback.
• Grounded, viewpoint-neutral output. The factual account is extractive: the excerpt's own words, the edit and technique findings, and, when the source is recovered, the full-context window of what came before and after. If the source is not found, it says so plainly. It flags rhetorical technique, never viewpoint, party, person, or the truth of a claim.
• Serves the suite. Omega Guard's orchestrator keeps the cheap edge signals (Content Credentials metadata, perceptual hashing) and delegates the trained-model verdict to Omega Lens.

What makes it different

It is sovereign and honest. The models run on your hardware with open weights, so nothing leaves the box, and the design is explicit that a verdict is corroborating evidence rather than proof. The context-restoration pass is deliberately nonpartisan: it exposes how a clip was cut and what surrounded it, without taking a side on the content.

Transparency and limits

State-of-the-art detectors are beaten by novel generators, so low-confidence and uncertain results must be treated as exactly that. Omega Lens produces scores and evidence for a human to weigh, and no verdict should drive an alert, a takedown, or an export without that human review.