Skip to content

VideoWorld

A generative model that learns from unlabelled video

Visit official site

What VideoWorld is

The model is described as a deep generative model that learns complex knowledge from unlabelled video through autoregressive generation.

What you can do with it

  • Learn task rules and operations from visual input
  • Represent multi-step visual change with an LDM
  • Generate the next frame from a previous one

Getting started

  • Download the open-source code and data
  • Turn video frames into discrete tokens with VQ-VAE
  • Train an autoregressive Transformer with next-frame prediction
  • Generate new frames and extract task operations

Pros & cons

✓ What we like

  • Open-source code and data are provided
  • The material states a 5-dan professional level in video Go

! What to watch out for

  • Learning starts from unlabelled video input
  • The documented pipeline uses discrete tokens from video frames

FAQ

What is the core model?

The material names a latent dynamics model, or LDM, for multi-step visual change.

What is released?

The feature list names open-source code and data.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review