AI Video Detector

This Actor estimates the likelihood that a video was generated using AI video generation models.

It uses forensic spatial and temporal signals instead of black-box deep learning classifiers.

The output is an explainable risk score (0–1) with a breakdown of contributing signals and a short interpretation.

---

🔍 How it works (high-level)

This Actor analyzes raw video frames and extracts low-level statistical signals that are commonly altered by AI video generation pipelines (diffusion / GAN-based systems).

The main signals are:

• Noise variance
  Real cameras contain natural sensor noise.
  AI videos are often over-smoothed and suppress this randomness.

• Temporal noise correlation
  In real videos, noise changes naturally from frame to frame.
  AI-generated videos often reuse or smooth noise across frames.

• High-frequency energy
  Real cameras preserve natural grain and fine detail.
  AI models often distort or hallucinate high-frequency components.

These signals are normalized and combined into an interpretable risk score.

⚠️ This Actor provides a probabilistic estimate, not a definitive classification.

---

📥 Input

Required input format

{
  "videoUrl": "https://commondatastorage.googleapis.com/gtv-videos-bucket/sample/BigBuckBunny.mp4",
  "fps": 20
}

📤 Output

Example output:

{ "risk": 0.183, "signals": { "noise_variance": 21.6843, "temporal_noise_correlation": 0.2453, "high_frequency_energy": 0.9043 }, "fps": 30, "frames_analyzed": 300, "input_type": "url", "interpretation": "Higher risk indicates stronger likelihood of AI generation. Scores between 0.3–0.6 indicate uncertainty." }

🐰 CGI and animation (important)

This Actor does NOT classify CGI or animated videos as AI-generated.

Examples that correctly score low:

Blender animations (e.g. Big Buck Bunny)

Pixar-style renders

Game engine cutscenes

CGI uses deterministic rendering pipelines, not neural frame synthesis.

A low AI score for CGI content is expected and correct behavior.

⚙️ Technical stack

Python 3.10

FFmpeg (frame extraction)

OpenCV

NumPy

Apify Python SDK

⏱️ Performance notes

Runtime scales with video length × FPS

Recommended video length: under 2–3 minutes

Higher FPS increases precision but also execution cost

⚠️ Limitations

Not trained on every AI video generation model

Heavy compression can reduce detection reliability

Results should be used as supporting forensic evidence, not absolute proof

🎯 Intended use cases

Media forensics

AI content research

Content moderation pipelines

Academic experiments

Explainable AI-detection workflows

🧠 Final note

This Actor is a signal-based forensic analysis tool, not a consumer-grade AI detector.

Transparency, explainability, and clearly defined limitations are intentional design choices.