NVIDIA open-sources LongLive-Plug: one distillation reusable across 54 video tasks

songhan_mit · x · 2026-10-01

NVIDIA released LongLive-Plug, an open-source 'once-for-all' distillation approach for video generation: train a plug-and-play acceleration LoRA once and reuse it across 54 downstream tasks — from world models and robotics to video editing — without re-distilling each model.

It distills three capabilities into separate LoRAs (few-step sampling, adjustable CFG via λcfg, long-context quality) that can be composed as needed. Works with any model based on MiniMax-H3 or Wan 2.2 5B/14B backbones.

Related event: NVIDIA Open-Sources LongLive-Plug for One-Shot Video Model Acceleration(3 posts)→

Original post →

More from Multimodal

Multimodal channel →