Reka inverse dynamics model

Reka’s inverse dynamics models estimate camera movement from video frames. Released on 2 October 2026, the code and weights use Apache 2.0.

Reka’s inverse dynamics models estimate how a camera moved between video frames. They work backwards from the observed pictures to a camera-motion estimate, which can help build controllable video and interactive world-model systems.

Reka released the models on 2 October 2026, with code and weights under Apache 2.0.

Two variants read the movement differently

Variant	Input approach	Trained parameter count
Flow model	Uses optical flow from a frozen RAFT component	1.8 million, excluding that frozen component
Pixel model	Works from image pixels	9.8 million

Optical flow describes how image features move between frames. The flow variant uses that representation to estimate camera movement; the pixel variant learns from the pictures directly. The model card gives the supported inputs and execution details.

The action is camera movement

This release concerns camera controls inferred from video. Those controls can supply training signals for an interactive world model, where a requested movement should produce a corresponding change in view.

Applications involving physical robot joints require their own action definitions and validation. For this release, assess camera-motion accuracy on the kinds of scenes and movement the application will encounter, using the exact variant and preprocessing documented by Reka.

Two variants read the movement differently

The action is camera movement

More in World Models