Free Deepfake Video Detector

Deepfake Video Detector — Is This Video Real or AI?

DeepFakeCheck's deepfake video detector analyzes face-swap videos, Sora-generated clips, and AI-edited footage. Upload an MP4, MOV, or WebM (up to 100 MB) and we'll sample the frames, run them through a dedicated face-deepfake model, and combine the result with a vision-language reasoning pass. The whole process typically completes in under 15 seconds — no signup, no watermarks, no upload limits.

How deepfake video detection works

Video is fundamentally different from image detection because a deepfake usually only affects part of the frame — a swapped face on a real body, or AI-synthesized lip movements on real audio. Our detection pipeline reflects that:

• Frame sampling. We extract up to 100 frames spaced across the video and run each through a face-detection model. Faces are cropped and passed to a dedicated deepfake classifier trained on FaceForensics++, DFDC, and recent Sora samples.
• Per-frame fake scoring. Each face crop gets a 0–100 "likely fake" score. We aggregate using the top-3 mean rather than the max, because a single bad frame is not enough to call a real video fake.
• Cross-modal sanity check. A vision-language model looks at the video as a whole — lip-sync alignment, scene consistency, lighting continuity, and whether the audio matches the apparent context.
• Weighted fusion. The frame model and reasoning model are combined with weights that depend on how many frames we sampled. Short clips get less weight on the frame model, because the sample is unreliable.

What types of deepfake videos we catch

We focus on the three categories that dominate real-world fraud and misinformation:

• Face-swap deepfakes (DeepFaceLab, Roop, Faceswap, commercial face-swap apps) — the swapped face usually has subtle boundary artifacts around the jawline and hairline.
• Fully AI-generated video (Sora, Runway Gen-3, Pika, Luma) — characterized by physically impossible motion, melted objects, and inconsistent shadows.
• Lip-sync deepfakes (Wav2Lip, HeyGen, Synthesia) — the mouth movements are AI-generated to match new audio, while the rest of the face is real.

Common use cases for the video detector

• Newsrooms verifying viral political clips before broadcasting.
• KYC and fraud teams spotting deepfake video during identity verification — a fast-growing attack vector in fintech.
• Influencer managers and brands monitoring whether the talent they represent is being deepfaked in scam videos.
• Family members checking suspicious video calls — voice-clone plus face-swap scams targeting older relatives are now common.

Video detection FAQ

How long can the video be? Up to 100 MB and roughly two minutes. For longer videos, trim to the most relevant segment — face-deepfake detection works best on clips where the suspected face is visible most of the time.

Does it work on heavily compressed clips from TikTok or X? Yes, but accuracy degrades. Re-encoded video destroys the high-frequency artifacts the face model uses; combine the result with your own reasoning.

Can it detect Sora-generated videos? Yes. Sora's failure modes — impossible physics, ghosted limbs, melted backgrounds — are explicitly part of our training signal.

What does "high risk" mean? It means our combined model is confident the video contains AI-generated content. It does NOT mean the video is necessarily harmful — it could be a legitimate creative use of Sora or a face-swap parody.

Is my video private? The video is uploaded to our processing pipeline over HTTPS, analyzed, and deleted within minutes. We never store, share, or train on uploads.