Skip to content
Try TracReal

Animations pause on keyboard focus, hidden tabs and offscreen sections. Reduced-motion preference shows static examples. Press Escape to pause all motion. Use each animation’s Resume control to restart it. Use the paper rail’s native scrolling and keyboard controls to explore. Automatic motion pauses while the rail has keyboard focus.

Detect what enters your organisation.Protect what leaves it.

Original photograph of people attending a video meetingUser-approved generated meeting illustration featuring a different woman
OriginalManipulated
Pioneered video-call deepfake detection.
World’s largest deepfake benchmarks owners
150+ research publications
7+ years of deepfake research
Multiple Best Paper Awards

Our products.

TracReal Detect

Detect manipulation. See the evidence.

Find AI-generated and manipulated images and video.

99%Detection accuracy
Loading demo media…

Inspect the face and flagged interval alongside the detection scores.

Review the highlighted face.

Inspect the face and flagged interval alongside the detection scores.

Audio
22/100
Video
87/100
Audiovisual
74/100

Flagged moment: 3.15–4.85 s.

Two generated videos and two edited images. Authored findings, not measured results; dataset public-use rights unverified, demo only.

Find areas to review

Highlight suspicious regions and moments.

Interpret the detection scores

Review the available image, video and audio signals.

Understand the result

Read a plain-language explanation alongside the findings.

Explore Detect ↗
TracReal Protect

Invisible protection for your images and videos.

Embed a watermark identifier that links your media to its registered source.

Orange trail shoe on wet black coastal rock, with breaking waves and a sunset
Illustrative linked-record concept: creator, organisation and contact details move into the photograph and return after a simulated watermark check. This represents linking and recovering a creator record, not a claim that these details are stored directly in the pixels. Original and protected use identical source pixels; no watermarking or ownership verification runs here.

Add a watermark without a visible logo

Embed an invisible identifier while preserving the media’s visual appearance.

Reconnect shared media to its source

Read the watermark identifier and retrieve its linked creator or rights-holder record.

Trace video from a single frame

Recover a watermark identifier from a single video frame.

Use TracReal where you work.

API

Detect and protect in code.

Bring both into your Python workflow.

Proposed Python SDK
from tracreal import TracReal
api = TracReal(api_key="key")
with open("media.mp4", "rb") as f:
    detected = api.detect(f)
    f.seek(0)
    protected = api.protect(
        f, identifier="id")
Planned interface · package forthcoming
Discuss API integration ↗
Web

Review in your browser.

Inspect detection and watermark findings.

Open TracReal ↗
Plugins

Check on your platform.

ChromeComing soon
XComing soon

Manipulated media. Real consequences.

The cost of trusting a familiar face without the full picture.

Our research

Explore Research ↗
Temporal localization

1M-Deepfakes Detection Challenge ↗

A challenge evaluates detection and localization of audio-visual manipulations, including small fake segments mixed into otherwise real videos, rather than whole-video labels alone.

2024 · ACM Multimedia challenge paper
Explainable detection

Explainable Deepfake Detection Challenge ↗

A challenge evaluates image classification alongside technical and accessible explanations, asking systems to identify manipulated entities and supporting evidence.

2026 · ACM Multimedia 2026 Grand Challenge; challenge paper/preprint
Multilingual detection

Tell me Habibi, is it Real or Fake? ↗

ArEnAV studies Arabic-English code-switching, dialect variation and monolingual Arabic content, extending audio-visual deepfake evaluation beyond predominantly monolingual benchmarks.

2026 · ICLR conference paper; published version of the 2025 preprint
Spatial and frequency cues

SFANet: Spatial-Frequency Attention Network for Deepfake Detection ↗

An ensemble combines transformer and texture approaches with frequency splitting, patch attention and face segmentation to study deepfake detection across varied datasets.

2025 · technical report/IEEE SPS Signal Processing Cup work; not a regular ICASSP proceedings paper
Let’s work with your media

See TracReal on your media.

Tell us what you need to detect, protect or integrate.

  1. 01Review your media.
  2. 02Explore the evidence.
  3. 03Plan your integration.

What would you like to explore?

Product interest
Phone number (optional)

For other countries, enter + country code and number.