Home How It Works Pricing Contact Start Free Trial
Home / How It Works
How It Works

From upload to verdict, in five deliberate stages.

Verafai isn't one model guessing "real or fake." It's a pipeline — purpose-built classifiers per modality, a scoring layer, a plain-language explainer, and a retraining loop that never stops learning from what it gets wrong.

Server infrastructure powering Verafai's inference pipeline
Live GPU Inference Cluster
The Pipeline

Click a stage to see what actually happens inside it.

Every file — video, audio, or image — passes through the same five stages. What changes underneath is which model does the heavy lifting.

01
Ingest

File or stream enters the system.

02
Analyze

Modality-specific model scans for artifacts.

03
Score

Confidence likelihood is computed.

04
Explain

Flag is translated into plain language.

05
Retrain

Result feeds the continuous training loop.

Under The Hood

The same five stages, three different specialists.

Video, audio, and images each fail in different, specific ways when they're synthetic. Verafai runs a dedicated classifier for each — not a single generic model spread thin.

Modality visual
Frame-by-frame video analysis — blending boundaries, lighting inconsistency, unnatural motion. VIDEO ARTIFACT CLASSIFIER

Video: Frame & Motion

Detects blending seams at face-swap boundaries, unnatural blink cadence, and lighting that doesn't match the scene — frame by frame, then aggregated across the clip.

CNN + Transformer

Audio: Spectral Signature

Voice-cloning tools leave behind pitch and prosody artifacts invisible to the ear but visible in a spectrogram. Verafai's classifier reads the frequency domain directly.

Spectral CNN

Image: Frequency Domain

Diffusion and GAN models leave a statistical fingerprint in the frequency domain that's invisible to the human eye but consistent across a generator family.

GAN Fingerprinting
Why It Doesn't Decay

Detectors go stale. Verafai's retraining loop is the whole point.

Every confirmed fake and every confirmed real sample gets fed back into training. When a new generator family appears, Verafai's target is to have a retrained, redeployed model within 14 days — not the months a static detector takes to fall behind.

01 · Collect

Every analysis — confirmed or disputed — becomes a labeled training sample.

+
02 · Retrain

Models are retrained on a rolling basis as new generator families are identified.

+
03 · Validate

Every retrained model is benchmarked against the full library of known generator families before it ships.

+
04 · Deploy

The updated model replaces the previous version in production with zero downtime.

+
Common Questions

Questions we get from every pilot team.

How accurate is Verafai, really?+
We're targeting greater than 95% accuracy on known generator families, benchmarked continuously as new models are added to our training set. Accuracy on brand-new, never-seen generator families is lower until the next retraining cycle closes the gap — which is why the 14-day retrain target matters as much as the headline number.
What file types and sizes are supported?+
Images (JPG, PNG, WEBP), video (MP4, MOV), and audio (WAV, MP3). Phase 1 supports files up to 500MB and clips up to 10 minutes; larger batch and archive processing arrives in Phase 3.
Can we run this via API instead of the web app?+
Yes — the REST API accepts programmatic uploads and returns a webhook callback on completion, designed to slot into KYC, claims, or moderation pipelines. Full API availability ships in Phase 2.
Do you store the media we upload?+
Analysis history and reports are stored per account for audit purposes. See our Privacy Policy for full detail on retention and deletion controls.
Get Access

Ready to see it work on your own media?

Join the pilot waitlist or request a live walkthrough with the founding team.