Skip to content

VideoPrism

A general video encoder for video understanding tasks

Visit official site

What VideoPrism is

The model is described as a general video encoder for video understanding tasks, built on a large and varied pre-training data set. The material also calls it a foundation model for video understanding.

What you can do with it

  • Classify video content with the encoder
  • Work on video localisation tasks
  • Retrieve videos from a collection
  • Generate captions and answer questions about video

Pre-training

  • Contrastive learning matches video and text in the first stage
  • Masked video patches are predicted in the second stage
  • The two stages draw on different supervision signals
  • The data set includes 36 million high-quality video-text pairs

Model notes

  • Video clips with noisy text are also part of pre-training
  • A fixed model adapts directly to downstream tasks
  • Results are reported on 30 video understanding benchmarks
  • The work is published on a Google Research page

Pros & cons

✓ What we like

  • A fixed model adapts directly to downstream tasks
  • Different supervision signals are used in both stages

! What to watch out for

  • Pre-training relies on a very large and varied data set
  • The noisy text clips are part of the pre-training corpus

FAQ

What tasks are named?

The material lists classification, localisation, retrieval, captioning and question answering.

How many benchmarks are mentioned?

The material reports results on 30 video understanding benchmarks.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review