README.md

metadata

license: apache-2.0
pipeline_tag: text-to-video
tags:
  - video-generation
  - sparse-attention
  - autoregressive

Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention

This repository contains the model checkpoints for Light Forcing, the first sparse attention solution specifically tailored for autoregressive (AR) video generation models.

Paper | GitHub Code

Introduction

Advanced autoregressive video generation models often suffer from the quadratic complexity of attention. Light Forcing addresses this bottleneck with two key innovations:

1. Chunk-Aware Growth: A mechanism to quantitatively estimate the contribution of each chunk, determining their sparsity allocation.
2. Hierarchical Sparse Attention: A strategy to capture historical and local context in a coarse-to-fine manner.

The method achieves a 1.2x–1.3x end-to-end speedup while maintaining high visual quality. When combined with FP8 quantization and LightVAE, it can achieve up to a 3.0x speedup on hardware like the RTX 5090.

Usage

For the full environment setup, the authors recommend using the provided Docker image:

docker pull lvchengtao/light_forcing:v1

Fast Inference

After setting up the environment and downloading the necessary checkpoints (e.g., Wan2.1-T2V-1.3B), you can run inference using the scripts provided in the repository.

For short-video generation (e.g., 5s):

python inference.py \
  --config_path configs/light_forcing_short.yaml \
  --output_folder videos/light_forcing_short \
  --checkpoint_path path/to/short_video_gen.pt \
  --data_path prompts/MovieGenVideoBench_extended.txt \
  --use_ema

For long-video generation (e.g., 15s):

python inference.py \
  --config_path configs/light_forcing_long.yaml \
  --output_folder videos/light_forcing_long \
  --checkpoint_path path/to/long_video_gen.pt \
  --data_path prompts/MovieGenVideoBench_extended.txt \
  --use_ema \
  --num_output_frames 63

Citation

If you find this work or the code useful, please cite:

@article{lv2026light,
  title={Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention},
  author={Lv, Chengtao and Shi, Yumeng and Huang, Yushi and Gong, Ruihao and Ren, Shen and Wang, Wenya},
  journal={arXiv preprint arXiv:2602.04789},
  year={2026}
}