ReWorld: An Interactive World Model with Long-Horizon Memory
TL;DR - ReWorld is an interactive video world model that combines real-time control with long-horizon visual memory under a fixed inference budget. It can stream 704Ă—1280 worlds while recalling and regenerating previously visited views during minute-long rollouts.
- Mixed local/global attention heads and randomized routing balance short-term action following with full-history learning.
- A bounded KV cache and pose-indexed landmark bank retrieve memories near the current camera pose without retaining full-history attention.
- Metric-aligned multi-source training and palindrome trajectories teach consistent physical controls and revisitation memory.
- LoRA-based distillation reduces generation to four sampling steps, while evaluations report leading control fidelity and video quality against six recent models.