SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.18636 · ICML · 2026

Attention Sparsity is Input-Stable: Training-Free Sparse Attention for Video Generation via Offline Sparsity Profiling and Online QK Co-Clustering

Chen Gao, Zhibo Chen, Cong Wang, Jiayu Chen, Qingyun Sun, Jianxin Li, Jiankun Wang, Jiayi Luo, Hanxin Zhu

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 11 functions out of this paper's own repositories and ran 8 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
Mutual-Luo/SVOO canonical 8 of 11
FunctionStatusWhere it lives
generate_temporal_head_mask_mod Ran Mutual-Luo/SVOO/svoo/models/hunyuan10/utils.py
code served (permissive licence) · get_code("680b148aa522eec6")
generate_temporal_head_mask_mod Ran Mutual-Luo/SVOO/svoo/models/wan/utils.py
code served (permissive licence) · get_code("f7ed76a6faefd4f2")
get_attention_mask Ran Mutual-Luo/SVOO/svoo/models/hunyuan10/utils.py
code served (permissive licence) · get_code("1e76f50cbc64301f")
hunyuan_token_reorder_to_frame_major Ran Mutual-Luo/SVOO/svoo/models/hunyuan10/placement.py
code served (permissive licence) · get_code("0543cdd0b9d08934")
hunyuan_token_reorder_to_token_major Ran Mutual-Luo/SVOO/svoo/models/hunyuan10/placement.py
code served (permissive licence) · get_code("0847ff97c9658511")
sparsity_to_width Ran Mutual-Luo/SVOO/svoo/models/wan/utils.py
code served (permissive licence) · get_code("793abbf309d92d09")
wan_token_reorder_to_frame_major Ran Mutual-Luo/SVOO/svoo/models/wan/placement.py
code served (permissive licence) · get_code("2cbbc31b189869b2")
wan_token_reorder_to_token_major Ran Mutual-Luo/SVOO/svoo/models/wan/placement.py
code served (permissive licence) · get_code("272c97b0d22d2805")
prepare_flashinfer_attention Not yet run Mutual-Luo/SVOO/svoo/models/wan/attention.py
code served (permissive licence) · get_code("2191dca04d00a3c3")
ref_hunyuan_sparse_head_placement Not yet run Mutual-Luo/SVOO/svoo/models/hunyuan10/placement.py
code served (permissive licence) · get_code("c5941af3e2811f0d")
ref_wan_sparse_head_placement Not yet run Mutual-Luo/SVOO/svoo/models/wan/placement.py
code served (permissive licence) · get_code("3c0438a386efef9e")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Diffusion Transformers (DiTs) achieve strong video generation quality but suffer from high inference cost due to dense 3D attention, leading to the development of sparse attention technologies to improve efficiency However, existing training-free sparse attention methods in video generation still face two unresolved limitations: ignoring layer heterogeneity in attention pruning and ignoring query-key coupling in block partitioning, which hinder a better quality-speedup trade-off. In this work, we uncover a critical insight that attention sparsity of each layer is its intrinsic property, with minor effects across different inputs. Motivated by this, we propose SVOO, a training-free Sparse attention framework for fast Video generation via Offline layerwise sparsity profiling and Online bidirectional co-clustering. Specifically, SVOO adopts a twostage paradigm: (i) offline layer-wise sensitivity profiling to derive intrinsic per-layer pruning levels, and (ii) online block-wise sparse attention via a bidirectional co-clustering algorithm. Extensive experiments on seven widely used video generation models demonstrate that SVOO achieves a superior quality-speedup trade-off over state-of-the-art methods, delivering up to 1.93× speedup while maintaining a PSNR of up to 29 dB on Wan2.1. Code is available at: https: //github.com/Mutual-Luo/SVOO.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2603.18636")
get_code_for_paper("2603.18636")
have("2603.18636")

Connect an agent — have() is free.