NVlabs/LongLive

What is it?

Key features

• 01Real-time interactive long video generation driven by sequential prompts (LongLive 1.0).
• 02NVFP4 and FP8 quantized training and inference, with sequence parallel support (LongLive 2.0).
• 03LoRA-based distillation that transfers capabilities across Wan2.1, Wan2.2 and MiniMax-H3 models (LongLive-Plug).
• 04Robot policy training and inference on LIBERO, RoboTwin 2.0 and RoboCasa benchmarks (Long-WAM).
• 05Each project has its own code, documentation and model weights, and can be fetched with sparse checkout.

Use cases

• Generating minute-long videos from a stream of text prompts for interactive creative tools.
• Serving long-video generation models on NVIDIA GPUs with quantization to reduce memory and latency.
• Reusing one distilled capability across several video generation backbones without retraining each.
• Training and evaluating robot manipulation policies on simulation benchmarks or real arms.

Quick start

1. 1

   Clone the repository with a partial, sparse checkout

   git clone --filter=blob:none --sparse --single-branch --branch main --depth 1 https://github.com/NVlabs/LongLive.git
2. 2

   Enter the cloned repository

   cd LongLive
3. 3

   Check out only the Long-WAM project

   git sparse-checkout set Long-WAM
4. 4

   Enter the Long-WAM project directory

   cd Long-WAM

Things to know

• !This is a monorepo of four separate research projects, so you must pick one and follow its own setup.
• !The benchmarked models were run on high-end NVIDIA GPUs such as H100 and GB200, so large GPU resources are likely needed.

Topics

FAQ

What is LongLive?

NVIDIA research monorepo with four projects for long video generation, world models and robot world-action models, each in its own directory. LongLive is a collection of NVIDIA research projects built around autoregressive video generation that runs for long durations. The repository is organized as four self-contained directories: LongLive 1.0 generates long videos in real time from sequential user prompts, LongLive 2.0 provides NVFP4 quantization and sequence parallelism for training and serving long-video models, LongLive-Plug distills reusable capabilities into LoRA adapters that attach to downstream video models, and Long-WAM provides long-context world-action models for training and deploying robot policies.

Is LongLive free to use? What license does it have?

LongLive is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.

Is LongLive still maintained?

Actively maintained. The last commit to LongLive was on Oct 19, 2026.

What can I use LongLive for?

Generating minute-long videos from a stream of text prompts for interactive creative tools. Serving long-video generation models on NVIDIA GPUs with quantization to reduce memory and latency. Reusing one distilled capability across several video generation backbones without retraining each. Training and evaluating robot manipulation policies on simulation benchmarks or real arms.

How do I install LongLive?

Clone the repository with a partial, sparse checkout: git clone --filter=blob:none --sparse --single-branch --branch main --depth 1 https://github.com/NVlabs/LongLive.git Enter the cloned repository: cd LongLive Check out only the Long-WAM project: git sparse-checkout set Long-WAM Enter the Long-WAM project directory: cd Long-WAM

Where can I get LongLive?

The source code and installation instructions are on GitHub: https://github.com/NVlabs/LongLive