Summary

ShadowDancer is a video world model research project that teaches a single model many different actions by learning unified dynamics representations from synchronized shadow pairs. It separates dynamics from appearance so an action can be transferred into new visual worlds.

Key Features

Learns unified dynamics representations from video and shadow pairs.
Disentangles action dynamics from visual appearance.
Supports action transfer into worlds the model has not seen.
Uses cross-shadow prediction as a core training signal.
Applies a block-causal world model for controllable rollouts.
Demonstrates human motion, combat, camera, and robot manipulation tasks.
Links to a public GitHub repository.
Includes multiple direct MP4 previews on the project page.

About ShadowDancer


The method constructs shadow pairs, performs cross-shadow prediction, and uses a block-causal world model to support action-controllable rollouts. Its examples cover human motion, combat-style movement, camera control, and robot manipulation, showing one latent interface across varied dynamics families.


ShadowDancer is useful for world-model researchers, controllable video generation teams, and simulation-oriented AI developers. It provides a framework for making video models respond to action commands without retraining a separate model for every domain.

Get more likes & reach the top of search results by adding this button on your site!

Embed button preview - Light theme
Embed button preview - Dark theme
TurboType Banner

Subscribe to the AI Search Newsletter

Get top updates in AI to your inbox every weekend. It's free!