NVIDIA · 2026-05-18 · notable
LongLive-2.0 — NVIDIA's NVFP4 Parallel Infrastructure Generates Minute-Long Interactive Video at 45.7 FPS
NVIDIA's follow-up to LongLive uses NVFP4 4-bit precision and sequence-parallel autoregressive training to speed up long video generation 2.15x in training and 1.84x in inference; the 5B model runs at 45.7 FPS.
A 4-bit training and inference stack that makes minute-long interactive video generation fast enough for real time.
Key specs
| GitHub stars | 1,253 |
|---|---|
| Inference speed | 45.7 FPS (5B model) |
| Training speedup | 2.15x |
| Inference speedup | 1.84x |
What is it?
LongLive-2.0 is an infrastructure update to NVIDIA's LongLive project for generating long, multi-shot, interactive videos. It targets the speed and memory bottlenecks that appear as generated videos get longer.
How does it work?
For training it introduces Balanced SP, a sequence-parallel autoregressive scheme that pairs clean-history and noisy-target temporal chunks on each GPU, combined with NVFP4 4-bit precision to cut memory and speed up matrix multiplies. Unlike Self-Forcing methods that need ODE initialization and distribution matching distillation, it tunes a diffusion model directly into an autoregressive one. For inference on Blackwell GPUs it uses W4A4 NVFP4 compute, an NVFP4-quantized KV cache, and asynchronous streaming VAE decoding.
Why does it matter?
It reports 2.15x faster training and 1.84x faster inference, with the LongLive-2.0-5B model hitting 45.7 FPS — fast enough for real-time, interactive long-video generation. Code, models, and demos are released under Apache 2.0.
Who is it for?
video-generation researchers
Try it
github.com/NVlabs/LongLive