Open family of causal diffusion representation video autoencoders that let interactive world models react to actions in under 230 ms on a single GPU, avoiding the latency of frame-packing used by typical text-to-video autoencoders.
CADRE is a family of causal video autoencoders developed for interactive world models. Its models explore different spatial and temporal video-compression settings, using frozen DINOv3 features and a causal diffusion decoder. Kyutai released the research paper, training and inference code, model weights, and an interactive demo. In the accompanying study, world models using CADRE were tested in human manipulation, robot manipulation, driving, and indoor navigation. The project examines how camera motion and compression choices affect responsiveness, motion quality, and action following.
Be the first Cosign
Add it to your Products I use list. You can also share why you use it.
