README.md

metadata

title: Light Forcing Video Generation
emoji: 🎬
colorFrom: gray
colorTo: pink
sdk: gradio
sdk_version: 6.15.1
app_file: app.py
short_description: Text-to-video with Light Forcing sparse attention
python_version: '3.12'
startup_duration_timeout: 30m

Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention

This Space demonstrates Light Forcing, a sparse attention acceleration method for autoregressive (AR) video generation models. It generates short videos (~5 seconds) from text prompts using the Wan2.1-T2V-1.3B model with Light Forcing's chunk-aware sparse attention.

How it works

Light Forcing introduces:

• Chunk-Aware Growth: quantitatively estimates the contribution of each chunk to determine sparsity allocation
• Hierarchical Sparse Attention: captures informative historical and local context in a coarse-to-fine manner

Model

• Base model: Wan-AI/Wan2.1-T2V-1.3B
• Light Forcing checkpoint: mack-williams/Light-Forcing
• Paper: arXiv:2602.04789
• Code: GitHub

License

Apache 2.0