Video Generation Demo

Generation examples

Scenes generated by Proposed method Discover the results

Showreel

A screenplay, brought to life

All background music in the showreel is generated by the model itself.

Examples

Qualitative results

Explore Proposed method generations.
Switch to comparisons for matched baselines.

Approach

Grounded screenplays

Persistent character identities and ordered events connect the screenplay
to the visible cast, their actions, and the soundtrack.

Speaker identity

Assign each line to a specific character in the frame.

Ordered actions

Specify who acts and when each action takes place.

Camera motion

Describe camera movements within the event sequence.

Audio events

Direct dialogue, ambient sound, and scene-specific effects.

A Director that sees the frame

Qwen3-VL reads the first frame and screenplay together. Its grounded features guide every audio and video block of LTX-2.3.

First frame + ScreenplayQwen3-VL DirectorJoint audio + video
Model architectureShow diagramHide diagram
Inside Proposed method Director · Audio · Video
Open full size ↗
Proposed method architecture: Qwen-VL encodes the first frame and screenplay, then audio and video adapters guide the joint generator alongside native Gemma text conditioning.

Zoom to inspect the connections. Scroll to explore; use + / − to zoom and 0 to fit.

Grounded Director features complement the backbone’s native text condition.
Loading the films…

Review supplement

Additional videos and tables, organized by reviewer and question.

Open supplementary results ↗