The Forensic Pipeline V2
Discover exactly how Truth Lenses detects digital manipulation - from client-side metadata extraction to high-speed multimodal vision verification.
Client Metadata & In-Memory Sampling
When you analyze media, it never touches persistent storage. The browser extracts EXIF camera metadata and inspects for C2PA provenance markers locally. Large images (>2048px) are smoothly downscaled on Canvas, while video clips are sampled locally into 8 key frames-raw video files never leave your device.
Forensic Parameterization
We supply a structured system prompt instructing the vision model to inspect visual clues in plain English: anatomical inconsistencies, lighting irregularities, texture smoothing, and temporal flickering across video frames-grounded by client metadata context.
Stateless Multi-Model Free Cascade
The media payload is analyzed statelessly in memory using a resilient priority cascade across Google Gemini models (Gemini 4 Hydrogen, Gemini 3.8 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite). The API checks structural patterns and discards all data immediately upon completion.
Rich Structured Verdict
The analysis returns a validated JSON schema containing a verdict (REAL, FAKE, ALTERED, UNCERTAIN), confidence percentage (0–100), categorized clue cards with severity ratings, file provenance context, and practical next steps.
1// Step 1: Client Metadata & In-Memory Sampling
const prepareMedia = async (file: File) => {
// Extract client EXIF & scan for C2PA provenance markers
const metadata = await extractClientMetadata(file);
if (file.type.startsWith('video/')) {
// Extract 8 frames locally in browser via Canvas
return { items: await extractFramesFromVideo(file, 8), metadata };
}
// Smoothly downscale large images (>2048px) on Canvas
const scaled = await downscaleIfNeeded(file, 2048);
return { items: [{ base64: scaled.base64, mimeType: scaled.mimeType }], metadata };
};How Truth Lenses Works
See how Truth Lenses samples media in your browser and runs stateless AI visual checks with zero file storage.
Verdicts, Reliability & Honest Limits
What each result means and how the pipeline stays accurate and private
What REAL & FAKE Mean
REAL indicates consistent photographic lighting, natural textures, and expected lens cues. FAKE indicates strong patterns of full AI generation or synthetic face-swapping.
What ALTERED & UNCERTAIN Mean
ALTERED flags genuine media with targeted edits, composites, or heavy retouching. UNCERTAIN appears when heavy compression or low detail prevents a clear call.
Browser-First Preprocessing
Oversized photos resize in-browser, EXIF and C2PA markers parse locally via exifr, and video files stay on-device while Canvas samples 8 preview frames.
Multi-Model Gemini Cascade
Requests start on gemini-4-hydrogen and automatically fail over across Gemini Flash models during demand spikes for fast, resilient results.
No Watermark or Filename Bias
Our prompt instructs the model not to treat corner blog logos, publisher stamps, or stripped WebP metadata as proof that an image is real.
Probabilistic Verification Only
No automated detector is 100% infallible. Always pair tool verdicts with reverse-image search and source checks for high-stakes decisions.
Technical & Privacy FAQs
Image vs. Video Pipeline at a Glance
Truth Lenses uses distinct browser-side preprocessing tailored to each media format before submitting payloads to the stateless Gemini vision cascade.
Image Forensic Checker
Single-frame high-resolution visual and metadata analysis
Video Deepfake Scanner
Multi-frame sequential analysis with on-device extraction
Zero Storage Guarantee on Both Pipelines
Whether scanning a photo or a video clip, neither raw media files nor sampled frames are ever saved to disk, stored in databases, or used to train AI models. In-memory payloads are permanently discarded immediately after your verdict is rendered.