←Home

    Endless Video, Made in Real Time

    How autoregressive video generators produce video chunk by chunk as it streams: distilling slow diffusion teachers into fast few-step students, scaling long-video training, teaching models to anticipate and remember, spending compute where it matters, and responding to users mid-stream.

    In order

    1 stories
    rainbow light swirl spinning on black
    rainbow light swirl spinning on black

    Rainbow light swirl spinning on black · Photo: Flyd

    Latent · Streaming video distillation

    How reward scores tell a fast video model where to focus its training

    Stream-R1 weights each training video, frame and region by a reward model's verdict, edging past its direct baseline on benchmarks at the same speed.

    arXiv preprint · 5 May 2026