Rather — World model benchmarks

Models, measured.

Matched inputs. Repeated runs. Evidence you can watch.

Compare predicted robot outcomes with a simulator reference, using the same starting scene and supplied actions.

See how generated worlds follow camera controls and retain a scene after turning back.

Browse by capability

Does it look right?

Does the world behave right?

Does it stay coherent?

Can you steer what happens?

Does it follow the prompt?

Does it remember the world?