Mixpeek Logo
    Schedule Demo

    Fake Video Detection

    A comprehensive pipeline to detect AI-generated and deepfake videos using multi-modal analysis of visual, audio, and motion cues.

    video
    1.5K runs
    Deploy Recipe

    "Videos that are likely or "

    from mixpeek import Mixpeek
    

    client = Mixpeek("your-api-key")
    

    # Create a collection for Fake Video Detection
    client.collections.create(
        feature_extractors=[
            "real-person-classification",
            "visual-artifact-detection",
            "lipsync-analysis",
            "gibberish-text-detection",
            "blink-abnormality-detection",
            "optical-flow-spike-detection",
            "lighting-inconsistency-detection"
        ]
    )

    Feature Extractors

    Real Person Classification

    Analyzes video frames using a CLIP model (ViT-L-14) to generate a 'visual realness' score, checking for stylistic coherence and human-like appearance.

    5K runs

    Visual Artifact Detection

    Leverages Gemini Pro to inspect video frames for common signs of AI generation, such as weird textures, blended objects, and other visual inconsistencies.

    5K runs

    Lipsync Analysis

    Compares visual lip movement to a Whisper-generated audio transcript to detect audio-visual synchronization issues common in deepfakes.

    5K runs

    Gibberish Text Detection

    Performs OCR on video frames and checks for nonsensical or 'gibberish' text, a common artifact in AI-generated scenes.

    3K runs

    Blink Abnormality Detection

    Analyzes eye blinking patterns for unnatural regularity or complete absence, a common tell-tale sign of a deepfake.

    4K runs

    Optical Flow Spike Detection

    Uses dense optical flow to detect unnatural spikes or inconsistencies in pixel movement between frames, indicating potential video manipulation.

    3K runs

    Lighting Inconsistency Detection

    Detects sudden, physically improbable changes in lighting across video frames, a common artifact in stitched-together or generated videos.

    2K runs

    Retrievers

    Hybrid KNN Search

    Combines vector similarity with keyword matching for improved retrieval accuracy

    325K runs

    Live Demo

    FakeCheck