From the source
Kyutai built CADRE, a family of open video autoencoders for interactive world models, and used them to study how a world model should compress video.
- The lab says world models usually borrow autoencoders from text-to-video generators, which compress time heavily and add lag that never mattered for offline clip generation.
- A typical 8× temporal compression delays every action by at least 400 ms at 20 frames per second, while cognitive neuroscience finds people notice their own movement is late beyond about 230 ms.
- CADRE reads each frame with a frozen DINOv3 image model, and a small trained bottleneck sets the compression, so family members differ only in how much they compress.
- Kyutai trained seven autoencoders and the same world model on each across manipulation and navigation, using 23 public video datasets totalling about 76k hours.
- It compared them with five baselines, including the autoencoders of Wan 2.2, Cosmos and LTX-2, and judged rollouts on image quality, motion quality and action following, with latency and frame rate timed on one H100.
- Kyutai released the paper, code, models and a live demo where users drive the world models from their own keyboard.