Merged generator checkpoint (recommended): the AR-trained base generator and DMD-distilled LoRA adapter are already merged, so inference only loads one generator_ckpt.
Base generator + LoRA checkpoint: the release code can also load the base generator first, attach LoRA modules, and then load the LoRA weights. This is useful for debugging or for users who want to inspect the adapter separately.

Use only one layout at a time. If you use the merged checkpoint, do not configure a separate lora_ckpt or adapter section, otherwise the LoRA adapter would be applied a second time.

Installation

bash

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:

bash

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

Download this checkpoint repository:

bash

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

Configure Inference

Edit configs/inference.yaml:

Option A: Merged Checkpoint (Recommended)

yaml

checkpoints:
  generator_ckpt: checkpoints/longlive2_5b/merged_generator.pt

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 merged_generator.pt with the actual merged checkpoint filename in this repository. If your local config was copied from a base+LoRA setup, remove checkpoints.lora_ckpt and the top-level adapter section before running inference.

Option B: Base Generator + LoRA

yaml

checkpoints:
  generator_ckpt: checkpoints/longlive2_5b/generator.pt
  lora_ckpt: checkpoints/longlive2_5b/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

This layout should reproduce the merged checkpoint behavior, but it keeps the adapter explicit at runtime.

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:

text

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:

text

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

Each JSON file contains:

json

{
  "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:

bash

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

For the merged checkpoint, standard inference only needs checkpoints.generator_ckpt.
For the base+LoRA layout, set both checkpoints.generator_ckpt and checkpoints.lora_ckpt, and keep the adapter section.
Do not mix the two layouts. A merged checkpoint should not be used together with lora_ckpt or adapter.
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.

Citation

Citation will be updated after the paper is released.

bibtex

@article{longlive2,
  title   = {LongLive2.0: An NVFP4 Parallel Infrastructure for Long Video Generation},
  author  = {TODO},
  journal = {TODO},
  year    = {2026}
}