AI Paper Atlas

Research map / Video / Video generation

Mixture of Distributions Matters: Dynamic Sparse Attention for Efficient Video Diffusion Transformers

Yuxi Liu, Yipeng Hu, Zekun Zhang, Kunze Jiang, Kun Yuan

ICML 2026 · Video › Video generation · approach: Efficient video diffusion models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use MOD-DiT sparse attention instead of full attentionHunyuanVideo-13B · VBench promptsspeedup: 1.00× -> 2.05×Improved
use MOD-DiT sparse attention instead of full attentionWan2.1-14B · VBench promptsspeedup: 1.00× -> 1.75×Improved
use MOD-DiT instead of full attentionHunyuanVideo-13B · VBench Subject ConsistencySubject Consistency: 0.9582 -> 0.9398Hurt
use MOD-DiT instead of SVGWan2.1-14B · VBench Subject ConsistencySubject Consistency: 0.7986 -> 0.9427Improved
use MOD-DiT instead of Radial attentionHunyuanVideo-13B · VBench promptsPSNR: 26.72 -> 27.73Improved
use MOD-DiT instead of Radial attentionHunyuanVideo-13B · VBench promptsSSIM: 0.885 -> 0.879Hurt
use static one-time pattern prediction instead of dynamic piecewise-linear tracking (ablation)MOD-DiT · VBenchgeneration qualityHurt
reduce full-attention warm-up steps m to 8 (ablation)MOD-DiT · VBenchspeed and qualityMixed
add full-attention warm-up before sparse attention (ablation)MOD-DiT · early denoising stepsqualityImproved
add MOD-DiT predictor overhead (sparsity conversion and solve kernel) (ablation)MOD-DiT · HunyuanVideo 512x832x81TFLOPS vs full attention: 0.0297 and 0.0002 vs 7.6127 TFLOPSNo clear effect
use custom CUDA least-squares kernel instead of torch.lstsq (ablation)MOD-DiT · solver timesolve time: 100x fasterImproved
mask 50% of blocks with MOD-DiTHunyuanVideo · HunyuanVideo 512x832x81attention TFLOPS: 7.6127 -> 5.7096Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy