Purdue University sealPurdue University
EcoAI Lab

PoCoAttn

Position–Content Collaborative Sparse Attention for Fast and High-Fidelity Video Generation

  • Jianwen Wu
  • Meiling Wang
  • Junkang Zhou
  • Yipeng Shen
  • Haoran You

Purdue UniversityEcoAI Lab

222426281.2×1.4×1.6×1.8×+2.3 dBPoCoAttnBetterPSNR (dB)DiT speedupWan2.1-T2V-14B 720p
  • PoCoAttn
  • Sol-Attn
  • PISA
  • SVG-EAR
  • SVG2
  • RadialAttn
PSNR and DiT speedup, Wan2.1-T2V-14B 720p, attention density in percent
MethodDensityDiT speedupPSNR (dB)
PoCoAttn25%1.40×27.81
PoCoAttn15%1.59×26.36
PoCoAttn10%1.73×25.07
Sol-Attn33.8%1.31×24.12
Sol-Attn24.7%1.37×23.67
Sol-Attn14.9%1.63×22.79
PISA30%1.19×24.42
PISA25%1.28×23.96
PISA15%1.50×23.03
SVG-EAR27%1.09×24.28
SVG-EAR14.4%1.24×24.10
SVG231.2%1.11×23.60
SVG223.9%1.19×23.25
SVG214.3%1.35×22.41
RadialAttn39.8%1.09×23.69
RadialAttn26.4%1.30×22.50

PoCoAttn achieves a better speed–fidelity Pareto frontier than existing sparse attention methods.

DiT speedup at 15% attention density, over FlashAttention-3 dense attention

2.02×
LTX-2.3-22B1080p text-to-video
1.59×
Wan2.1-T2V-14B720p text-to-video
1.60×
Wan2.1-I2V-14B720p image-to-video
1.79×
HunyuanVideo-13B720p text-to-video
1.72×
Bernini720p video editing

Abstract

Video diffusion transformers rely on 3D full attention over tens of thousands of spatiotemporal tokens, and its quadratic cost dominates inference latency. Sparse attention reduces this cost, but existing methods face a dilemma: static methods apply predefined masks and cannot capture input-dependent attention, whereas dynamic methods reconstruct the sparse mask repeatedly, paying online to rediscover positional structure that is largely stable across inputs.

We propose PoCoAttn, a training-free sparse attention framework that separates input-stable from input-dependent attention and resolves the former offline and the latter online. Across four video diffusion models and three tasks, PoCoAttn consistently achieves a better speed–fidelity Pareto frontier than state-of-the-art (SOTA) methods, improving PSNR by 1.9 to 5.0 dB over baselines and delivering up to 4.1× attention and 2.0× denoising speedups over dense FlashAttention-3 attention.

Method

The Proposed PoCoAttn Framework

Citation

BibTeX

PoCoAttn
@article{wu2026pocoattn,
  title   = {PoCoAttn: Position--Content Collaborative Sparse Attention
             for Fast and High-Fidelity Video Generation},
  author  = {Wu, Jianwen and Wang, Meiling and Zhou, Junkang and Shen, Yipeng and You, Haoran},
  journal = {arXiv preprint arXiv:2611.20307},
  year    = {2026}
}