AI-Generated Video Frames For Synthetic Video Detection

A collection of video frames from AI-generated videos and real sources for training synthetic video detectors. The dataset includes frames from seven AI generation algorithms and real video frames from the VideoASID dataset. It was created by author ductai199x and uploaded to Hugging Face in August 2024.

Use Cases

• Train a binary classifier to distinguish real video frames from AI-generated frames using labels for each source algorithm.
• Benchmark detection models across different AI video generators like Sora, Pika, and Stable Video Diffusion.
• Analyze visual artifacts or features specific to certain generation algorithms (e.g., CogVideo vs. VideoCrafter v2).

Strengths

• Includes frames from seven distinct AI video generation algorithms, providing algorithmic diversity.
• Combines synthetic frames with real video frames from the VideoASID dataset for a labeled detection task.

Limitations

• The real video component is incomplete, as frames from the Moment-in-Time dataset are excluded due to licensing.
• Dataset size, row count, and exact composition are unspecified in the provided metadata.

Provenance

Source

AI-generated frames from specified algorithms; real frames from VideoASID and (referenced) Moment-in-Time datasets.

Collection Method

Curated collection of frames from various AI video generators and public real video datasets.

Freshness

Last updated in August 2024.