Tracer Vision compares unknown images and videos against a reference corpus you control—using a local visual-similarity model (SigLIP) to surface material that hash matching misses. It runs fully offline and exports Project VIC 2.0 for Griffeye.
Known-file hashing catches what's already catalogued. But altered, re-cropped, re-compressed, and never-before-seen material slips straight through—and your analysts have no fast way to know which of thousands of unknowns to look at first.
MD5 and SHA-1 are bit-for-bit. Change a single pixel, re-encode, or re-crop and the hash changes—so the file reads as brand new even when it's the same scene you've seen a hundred times.
Faced with a folder of unknowns, analysts have no signal for where to start. Reviewing everything by eye is slow and keeps people in front of harmful material longer than necessary.
Many similarity tools require uploading the very material you're investigating. For this work, sending it to someone else's server is a non-starter.
Every check run gives each unknown up to three independent signals—visual similarity, hash matches, and review priority. You read them together; none overrides the others.
The primary signal. Each unknown is embedded locally with Google's SigLIP encoder and compared to your seed corpus by squared-Euclidean distance. Perceptual matching catches variants, crops, and re-compressions that hashing never will—reported as a tunable similarity band.
Import one or more Project VIC OData datasets and every unknown's MD5 and SHA-1 is looked up against them. Hits appear as chips next to the SigLIP score—they never change or override the visual-similarity result.
Videos are sampled frame-by-frame with ffmpeg, each frame embedded, and the closest-matching frame reported as the finding. Per-frame embeddings are cached, so re-checking the same video later never needs ffmpeg again.
One click produces a Project VIC 2.0 JSON report (or CSV) ready to ingest into Griffeye GID and your existing case management. Tracer Vision feeds your toolchain instead of replacing it.
A single binary that runs on a laptop—no server, no account, no per-image cloud calls. No telemetry, no auto-update beacon. Seeds are never rendered; unknowns stay masked behind a click-to-reveal so analysts opt in to view each file.
Optional and opt-in per run: with an API key, each unknown can be classified by Google's Content Safety endpoint into a review priority (VERY_LOW…VERY_HIGH). Only the embedding vector is sent—never the image. The one part of the workflow that touches the network, and it's off by default.
Tracer Vision sits at the discovery step. Curate what you're looking for, point it at a pile of unknowns, and it tells you which items most resemble your reference set—so analysts review the right things first.
Point it at folders of reference images. Tracer Vision computes and stores embeddings and hashes—no source paths, no thumbnails. The originals can leave the host afterward.
Drop a folder of unknown images and videos. Each is embedded locally and compared to every seed; the closest match is reported with a similarity band, VIC hits, and an optional review priority.
A results table ranks unknowns by similarity, with hash chips and priority side by side. History persists every check so you can re-match against an updated corpus or find similar unknowns.
Produce a Project VIC 2.0 JSON (or CSV) report and ingest it into Griffeye GID. The findings flow into the toolchain and case management your team already runs.
Most analysts live in the desktop GUI—Seeds, Check, History, and Settings views with drag-and-drop, masked previews, and one-click export. For labs that script their pipeline, the same engine drives a full CLI. Both produce identical Project VIC output.
# Build a reference corpus $ tracer-vision ingest --dir ./seeds --kind seed # Check a folder of unknowns, write a report $ tracer-vision match --dir ./unknowns --out report.json # Load a Project VIC dataset (additive hash signal) $ tracer-vision vic-import --file vic.json --source NCMEC # …or just open the desktop app $ tracer-vision gui
Tracer Vision points your analyst at the items most likely to be worth a closer look first. It runs on material you already possess. Every signal it reports—visual similarity, hash matches, review priority—is a hint; an analyst always makes the call.
Tracer Vision is licensed free of charge to qualifying U.S. law enforcement—solely for investigating, deterring, or prosecuting crimes against children. Tell us about your agency and we'll get you set up.
We'll be in touch shortly.