Sparse VideoGen & Quant VideoGen

Towards ultra-efficient video generation

Sparse VideoGen 1
Identity the spatial and temporal sparsity patterns in video diffusion models.
Access SVG1
Sparse VideoGen 2
Densify attention's critical tokens with semantic-aware permutation.
Access SVG2
Sparse VideoGen EAR
Mitigate sparse attention errors with error-aware routing and linear compensation.
Access SVG-EAR
Quant VideoGen
Auto-regressive long video generation via 2-bit KV-cache quantization.
Access QVG

Quick Links

SVG1 arxiv SVG2 arxiv SVG-EAR arxiv QVG arxiv Github Repo