What it is
An 8B diffusion world model for multi-turn video generation that follows free-form text and camera controls while carrying forward scene history across segments. Instead of treating all past prompts equally, it separates persistent scene context from time-local action commands via action-aware causal attention, then uses a VLM prompt planner to turn multimodal history plus user input into the next segment condition.
Key results
- Beats existing world models in human evaluation on interactive tasks
- Supports minute-scale rollouts with multi-turn text and camera control
- Uses an 8B diffusion backbone with packed sequences up to 360K tokens
Structured captions plus temporally scoped action conditioning are key to preserving world state while still obeying new interactive commands.