Video-model developers can explore more consistent object and scene reappearances without retraining the entire generator. Core capabilities include Spatiotemporal KV selection, Historical detail retrieval, Frozen generator backbone, and Lightweight trainable router.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Fixed active-memory budget, Text-to-video experiments, Image-to-video experiments, and Consistency across scene revisits.
