CADRE
A family of open causal diffusion video autoencoder models comprising seven variants, designed for interactive world models. It employs causal temporal modeling and diffusion mechanisms and supports low-latency streaming video generation.
Timeline (last 90 days)
Kyutai released CADRE, a family of open causal diffusion video autoencoders comprising 7 models.
According to MarkTechPost, a world model built on CADRE can run at 20 fps on a single H100.
According to MarkTechPost, the end-to-end reaction latency of the world model built on CADRE is below 230 ms.
CADRE places the diffusion mechanism inside the autoencoder rather than using diffusion only outside the latent space.
CADRE is not a complete world model but a family of video autoencoders; the dynamics prediction part of the world model must be built separately.
Public information lacks quantitative comparisons of CADRE's generation quality, the specific differences among the 7 models, and performance at different resolutions.
CADRE's target scenario is interactive world models, with a focus on reducing end-to-end latency and maintaining streaming generation capability.
CADRE can run on a single H100 and does not require multi-GPU parallelism or a cluster.
All Facts (8)
CADRE can run on a single H100 and does not require multi-GPU parallelism or a cluster.
50%UnverifiedAccording to MarkTechPost, a world model built on CADRE can run at 20 fps on a single H100.
50%UnverifiedAccording to MarkTechPost, the end-to-end reaction latency of the world model built on CADRE is below 230 ms.
50%UnverifiedCADRE places the diffusion mechanism inside the autoencoder rather than using diffusion only outside the latent space.
50%UnverifiedCADRE is not a complete world model but a family of video autoencoders; the dynamics prediction part of the world model must be built separately.
50%UnverifiedPublic information lacks quantitative comparisons of CADRE's generation quality, the specific differences among the 7 models, and performance at different resolutions.
50%UnverifiedCADRE's target scenario is interactive world models, with a focus on reducing end-to-end latency and maintaining streaming generation capability.
50%UnverifiedKyutai released CADRE, a family of open causal diffusion video autoencoders comprising 7 models.
50%