Deep Scan
The vault
that looks out for you.
Deep Scan sweeps public platforms for synthetic and unauthorised use of a performer's likeness — deepfakes, AI “concept trailers”, reposted scans — and matches what it finds against the one thing no generic detection service holds: the performer's ground-truth capture data, already in the vault.
Review queue
3 hits awaiting review · last sweep 2h ago
Instagram Reels
@velmirastudiosMediumNew“MARLOWE QUINN returns — TIDEWATER 2 (2027) Concept Trailer #tidewater #marlowequinn Watch the dark What If story where…”
Likeness match
72%
AI-generated
75%
Also in this content
3 match signals · adjudicator note
“Marlowe Quinn AI set — full renders in comments”
Likeness match
91%
AI-generated
88%
4 match signals · adjudicator note
YouTube Shorts
@fancutsHDMediumTakedown requested“Tidewater 2 — AI fan trailer (Marlowe Quinn, 4K)”
Likeness match
84%
AI-generated
90%
Also in this content
2 match signals · adjudicator note
The launch trailer
A hundred seconds · tap to pause
Vault-anchored
Detection compares against the canonical scan — the most current, highest-fidelity record of a performer that exists.
Five independent signals
Identity, synthesis, derivation, body geometry — and a fifth that stays ours.
Event-aware
Sweeps surge around cast announcements and trailer drops, when synthetic waves actually arrive.
Evidence-grade
Every flag carries specific, reviewable rationale — ready for a takedown letter.
The ground-truth advantage
Everyone can look.
Only the vault can compare.
A convincing fake can be trained from anything — press photos, red-carpet footage, a leaked still, or the scan data itself. Generic monitoring services can only compare suspect content against the same public imagery everyone else can see. ImageVault matches against the performer's own archived captures: calibrated scan stills, mesh geometry, and the exact files every production uses. When the source of a fake is scan-grade data, the party holding the original scans is the only one who can prove it — and as generators improve and visual artifacts disappear, detection built on possession of the source is the approach that keeps working.
Every scan a performer archives makes their monitoring measurably stronger. Storage and protection stop being separate products.
Months ahead of the public material
Deepfakes get built from posted trailers and lagging set leaks. ImageVault sits inside the production workflow, so it holds the most current, detailed record of the artist from the day of capture — months before any of that material is public.
A reference gallery built from real captures
Archived scan packages give the matcher the artist's face from every angle — the same multi-angle, studio-calibrated coverage a production needs for reference is exactly what face-matching needs. The real person, not a fan-site crawl.
Turntable renders unlock mesh-only packages
Packages with no photographic stills aren't blind spots: the pipeline renders reference stills from the scan's own 360° footage or a lit turntable of the mesh itself — coverage generated from geometry only the vault holds.
A derivation index over source imagery
Every reference still is perceptually fingerprinted. Reposts, leaks, screenshots, and re-renders of vault imagery match by hash — robust to recompression and resizing.
How detection is layered
Five signals, each proving something different
No single detector survives generator progress, so the monitor never relies on one. Each candidate is scored by independent layers — and a flag requires evidence of likeness and evidence of synthesis or derivation, never one alone.
01
Identity
Proves: the person shown is the talent
Face matching against the vault-anchored reference gallery — the performer's own scan stills, not public photos. Identity evidence never decays.
02
Synthesis
Proves: the media is AI-generated
Embedded provenance markers (C2PA / declared-AI metadata) checked first — deterministic and near-conclusive. A vision model then looks for generation artifacts, with generator-family attribution.
03
Vault-onlyDerivation
Proves: it was built from vault imagery
Perceptual-hash matching against fingerprinted reference stills catches reposts, leaks, screenshots, and re-renders of the source imagery — imagery nobody else holds.
04
Vault-onlyBody geometry
Proves: the body shown is — or isn't — the artist
The scan's mesh carries the artist's true proportions at production grade. Geometric ground truth no public photo holds, anchoring full-body likeness claims.
05
ProprietaryUndisclosed
Proves: what we keep to ourselves
A fifth, proprietary signal rides the platform's own workflow. Describing how it works publicly would help exactly the people it catches — so we don't.
Precision by design
A likeness match alone never flags — press photos and fan edits would drown the queue. Synthesis evidence alone never flags either — AI content that isn't the performer is not their problem. Only the combination raises a hit, and an unmeasured signal is recorded as exactly that: not measured, never “low”, and never presented as evidence of authenticity.
Built for where generators are going
Artifact detection is an edge that narrows with every generator release — which is why it's one layer here, not the strategy. Identity anchored to ground-truth captures, derivation matched against source imagery, and signals embedded where only the platform can put them don't decay as fakes get better. That's the part only a vault can do.
The coverage loop
Every scan you archive strengthens detection
Detection coverage is scored per talent, directly from what their vault contains. Archive a new scan package and its stills join the reference gallery and the derivation index automatically; even mesh-only packages contribute through pipeline-rendered turntable stills. The tier is honest about what it measures — reference quality, with concrete next-upload suggestions to climb it.
- Reference gallery synced automatically from archived packages
- Mesh and video-only scans covered via derived reference stills
- Next-upload suggestions show exactly what would strengthen matching
Unanchored
12/100No usable reference imagery yet. Sweeps still run on name and text intent.
Baseline
40/100First reference stills indexed from an archived scan package.
Anchored
72/100A working reference gallery — mesh-only packages reach this via derived turntable renders.
Fortified
100/100Photographic diversity across angles and sessions. The strongest reference set the vault can build.
How a sweep works
From public platform to reviewable evidence
01
Discover
Live sweeps crawl nine public platforms — short-form video, social feeds, Reddit, image search, stock libraries, and AI-generation sites — on a per-talent cadence, with queries and watched accounts tuned to each performer.
02
Score
Every candidate is scored by the five detection layers against the vault's reference set and derivation index — plus signals only the platform can carry.
03
Adjudicate
An AI adjudicator weighs the signals with full context — active vigilance windows, press-material priors, body-geometry context — and writes a specific rationale.
04
Act
Confirmed hits flow into graduated outreach — from a licence offer to a formal takedown — with the evidence trail attached. Every human verdict feeds back into detector calibration.
Account watchlist
Ordered by priority
@velmirastudios
TikTokOn watchlistVelmira Studios · 155k followers
80
Priority
4.9M views across flagged posts · active this week · 4 open hits
Reach
4.9M
views on flagged posts
Posts
4
4 open
First seen
6d ago
Last seen
6d ago
Also targeting 2 other protected talent on ImageVault
A pattern across multiple represented people is evidence of a commercial operation — it escalates to platform partner channels rather than a per-post report.
@velmirastudios
InstagramReported to platformCross-platform sibling · confirmed by matching captions
64
Priority
Hits roll up into an account watchlist — reach-ranked operators, not a flat list of posts.
What Deep Scan does
A two-minute tour · tap to pause
Capabilities
Monitoring that behaves like an investigator
Vigilance windows
Synthetic content arrives in waves triggered by cast announcements and trailer drops — often tagged with the character, not the actor. A vigilance window adds persona and production vocabulary to discovery for a bounded period, so a fake that never names the performer is still found.
Cross-platform pursuit
Misuse operators build audiences, not single posts. When an account is flagged, the monitor probes for the same operator on other platforms — and only confirms a sibling when its content actually matches, never on the name alone.
Account watchlists
Accounts that have hit once are harvested on every subsequent sweep, with reach tracked — so enforcement priorities follow audience size, not posting order.
Graduated outreach
Not every hit deserves a legal letter. Outreach templates run from warm (a licensing offer) to cold (formal takedown), each pre-filled with the hit's specific evidence and rationale.
Human-verdict feedback loop
Every confirmation, dismissal, and whitelist decision is read back as a calibration signal — per-detector and per-talent — so the system learns where it was over- or under-confident.
Audit and evidence
Sweeps, adjudications, and every hit's signal readings are recorded. Evidence trails say why something was flagged — specific observations, ready for enforcement.
Nine-platform coverage
Instagram Reels, TikTok, YouTube Shorts, X (Twitter), Pinterest, Reddit, Google Images, Getty / Shutterstock, and AI-generation platforms — each toggleable per platform, so coverage expands deliberately rather than by default.
Adult communities, badged
Adult communities are exactly where likeness misuse concentrates, so they're inside the sweep — and flagged hits carry an NSFW badge, warning talent before they tap through.
Managed by your representative
Talent never have to watch the queue alone — representatives triage hits, send takedowns, and manage watchlists on their clients' behalf, with every action on the record.
The fakes are coming either way.
Hold the original.
Detection anchored to ground-truth capture data is only possible for the platform that holds it. Archive the scan; the monitoring comes with it.