LongLive

// Overview

LongLive is NVIDIA Research's open-source stack for long, interactive video generation. LongLive 1.0 (ICLR 2026) established the interactive half: it accepts a sequence of user prompts and keeps generating video in real time as they arrive, using attention sink, KV-recache and streaming long tuning to hold quality across a long rollout. That version now lives on the repository's `v1.0` branch with its training and inference code and the LongLive-1.3B weights.

LongLive 2.0 is the infrastructure release: an NVFP4 parallel stack covering both training and inference. On the training side it adds balanced sequence parallelism for text-to-video and image-to-video autoregressive teacher-forcing, multi-shot (or single-shot) training, and NVFP4 or BF16 for both AR training and few-step distillation. On the inference side it does W4A4 NVFP4 with an NVFP4 KV cache, TorchAO FP8 W8A8 post-training quantisation from the BF16 checkpoint, multi-shot attention sink, sequence-parallel inference and async decoding.

The repository is maintained as working infrastructure rather than a frozen paper drop: FP8 inference landed in July 2026, the NVFP4 inference path was tuned with fused Triton RoPE/adaLN kernels, reduced KV-cache synchronisation, in-place quantised KV-cache updates and pinned VAE transfers for an 18.6% throughput gain, and I2V teacher-forcing and DMD distillation for Wan2.2-TI2V-5B arrived alongside. It carries a full documentation site covering installation, NVFP4 setup, training modes, inference and data organisation, plus a companion retrieval-augmented project, LongLive-RAG. In October 2026 the repository also gained Long-WAM, a world-action model for robot control built on a LongLive 2.0 Robot video model, with evaluation scripts for LIBERO, RoboTwin 2.0, DOMINO and RoboCasa and real-robot interfaces for YAM, Franka and Unitree G1.

// What it does

• Real-time interactive generation from a stream of user prompts (LongLive 1.0: attention sink, KV-recache, streaming long tuning)
• NVFP4 W4A4 inference with an NVFP4 KV cache, plus TorchAO FP8 W8A8 post-training quantisation from the BF16 checkpoint
• Balanced sequence parallelism for T2V and I2V autoregressive teacher-forcing training, single- or multi-shot
• NVFP4 or BF16 for both AR training and few-step distillation
• Multi-shot attention sink, sequence-parallel inference and async decoding
• Full documentation site covering installation, NVFP4 setup, training modes, inference and training-data layout

// Getting started

Installation, NVFP4 setup, training and data layout are documented on the project's docs site; the commands below are the repository's own quick start. The validated FP8 stack is Python 3.10, PyTorch 2.8.0+cu128 and TorchAO 0.13.0 on H100, and compute capability 8.9 or newer is required.

git clone --single-branch --branch main --depth 1 https://github.com/NVlabs/LongLive.git

import torch
from omegaconf import OmegaConf

from pipeline import CausalDiffusionInferencePipeline
from utils.config import normalize_config
from utils.inference_utils import (
    load_generator_checkpoint,
    place_vae_for_streaming,
    prepare_single_prompt_inputs,
    save_video,
)

prompt = "A compact silver robot walks through a clean robotics lab."
config = normalize_config(OmegaConf.load("configs/inference.yaml"))
device = torch.device("cuda")

torch.set_grad_enabled(False)
pipe = CausalDiffusionInferencePipeline(config, device=device)
load_generator_checkpoint(pipe.generator, "LongLive-2.0-5B/model_bf16.pt")
pipe = pipe.to(device=device, dtype=torch.bfloat16)
place_vae_for_streaming(pipe, config)
pipe.generator.model.eval().requires_grad_(False)

noise, prompts = prepare_single_prompt_inputs(config, prompt, device)
video = pipe.inference(noise=noise, text_prompts=prompts)
save_video(video[0], "videos/quickstart/sample.mp4", fps=24)

python inference.py --config_path configs/fp8/inference_fp8.yaml

configs/nvfp4/inference_nvfp4.yaml

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Generate minute-scale video that responds to prompts as they arrive, rather than rendering a fixed clip
• Train or distil your own autoregressive video model with sequence parallelism and low-precision support already wired in
• Cut inference cost on a 5B video model with NVFP4 W4A4 or FP8 W8A8 quantisation
• Study a worked implementation of attention sink, KV-recache and few-step distillation for long video

// Version history

Every verified update to LongLive that AI/TLDR tracked, newest first — each links to our coverage and the official changeset.

1. 2026-10-07

   Long-WAM landed in the repository's Long-WAM folder: a world-action model for robots that pretrains a LongLive 2.0 Robot video model, then adapts it to predict actions using up to 19.2 seconds of visual history. Code and checkpoints are Apache 2.0.

   our coverage →changeset ↗
2. 2026-05-182.0

   LongLive 2.0 shipped as an NVFP4 parallel infrastructure for long video generation, covering AR training, DMD distillation and inference; NVIDIA reported 2.15× faster training, 1.84× faster inference and 45.7 FPS for the 5B model.

   our coverage →changeset ↗

// How LongLive compares

LongLive alongside other open-source video world models tools AI/TLDR tracks, ranked by GitHub stars.