LongLive-2.0-5B

<p align="center"> <img src="logo.png" alt="LongLive2.0 logo" width="100%"> </p>

LongLive2.0 5B

This repository hosts temporary LongLive2.0 5B BF16 checkpoints for inference with the LongLive2.0 release code:

https://github.com/NVlabs/LongLive

The checkpoint package contains two parts:

• Base generator checkpoint: the AR-trained Wan2.2-TI2V-5B generator.
• LoRA checkpoint: the DMD-distilled few-step LoRA adapter.

LongLive2.0 inference loads the base generator first, applies the LoRA modules, and then loads the LoRA weights.

Installation

git clone https://github.com/wileewang/LongLive2.0.git
cd LongLive2.0

conda create -n longlive2 python=3.10 -y
conda activate longlive2
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

The released LongLive2.0 checkpoint is sufficient for standard inference. You only need to download the original Wan2.2-TI2V-5B components if you want to run training, initialize from the original Wan weights, or use code paths that explicitly load the base Wan model files:

huggingface-cli download Wan-AI/Wan2.2-TI2V-5B \
  --local-dir wan_models/Wan2.2-TI2V-5B

Download this checkpoint repository:

huggingface-cli download Perflow-Shuai/longlive_2.0_5B_tmp_20260507 \
  --local-dir checkpoints/longlive2_5b

Configure Inference

Edit configs/inference.yaml:

checkpoints:
  generator_ckpt: checkpoints/longlive2_5b/path/to/base_generator.pt
  lora_ckpt: checkpoints/longlive2_5b/path/to/dmd_lora.pt

adapter:
  type: lora
  rank: 128
  alpha: 128
  dropout: 0.0
  verbose: true

data:
  data_path: /path/to/inference_prompts

output_folder: videos/longlive2
num_samples: 1

inference:
  sampling_steps: 4
  sink_size: 8
  guidance_scale: 1.0
  multi_shot_sink: true
  multi_shot_rope_offset: 8

Replace the checkpoint filenames above with the actual files in this repository. If the LoRA checkpoint is not used, remove the adapter section and leave lora_ckpt unset.

Prompt Folder

data.data_path is passed to MultiTextConcatDataset in inference.py. It can be either:

• a .txt file, where each line is one single-shot prompt; or
• a directory of multi-shot prompt folders.

For a directory input, the code supports both of the following layouts. The direct caption-root layout is the simplest:

inference_prompts/
  robot_lab_demo/
    0.json
    1.json
    2.json
    shot_durations.txt

It also supports a dataset root with an outer caption/ folder:

inference_prompts/
  caption/
    robot_lab_demo/
      0.json
      1.json
      2.json
      shot_durations.txt

Each JSON file contains:

{
  "caption": "A compact silver robot with one blue optic explores a clean robotics lab."
}

shot_durations.txt is optional. If provided, each number is the number of temporal chunks assigned to the corresponding caption, for example:

2 2 4

Run

Single node, 8 GPUs:

torchrun --standalone --nnodes=1 --nproc_per_node=8 inference.py \
  --config_path configs/inference.yaml

Single GPU:

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

Outputs are written to output_folder.

Notes

• The base checkpoint and LoRA checkpoint should be loaded together for the few-step DMD model.
• inference.sampling_steps controls the number of denoising steps.
• inference.multi_shot_sink enables the multi-shot attention sink.
• inference.multi_shot_rope_offset controls the multi-shot RoPE offset.
• For NVFP4 inference, use the separate NVFP4 config and setup instructions in the LongLive2.0 documentation.

License/Terms of Use

GOVERNING TERMS: This trial service is governed by the NVIDIA API Trial Terms of Service. Use of this model is governed by the NVIDIA Open Model License Agreement.

Citation

@article{longlive_2,
  title={LongLive2.0: An NVFP4 Parallel Infrastructure for Long Video Generation},
  author={Chen, Yukang and Wang, Luozhou and Huang, Wei and Yang, Shuai and Zhang, Bohan and Xiao, Yicheng and Chu, Ruihang and Mao, Weian and Hu, Qixin and Liu, Shaoteng and Zhao, Yuyang and Mao, Huizi and Chen, Ying-Cong and Xie, Enze and Qi, Xiaojuan and Han, Song},
  journal={arXiv preprint arXiv},
  year={2026}
}