arXiv AI

Temporal-Attention Head Specialization During Video Diffusion Training

arXiv AI
Oct 31

Focused Forcing: Content-Aware Per-Frame KV Selection for Efficient Autoregressive Video Diffusion

Focused Forcing is a training‑free method for selecting key‑value (KV) cache entries in autoregressive video diffusion models. It selects historical frames per generated frame by combining attention and diversity scores, and allocates larger budgets to more important heads. The approach yields up to 1.48× faster generation while improving visual quality and text alignment across several generation paradigms.

By Peiliang Cai, Evelyn Zhang, Jiacheng Liu, Hao Lin, Ruiqi Zhang, Weile Mo, Yue Ma, Shikang Zheng, Jiehang Huang, Dongrui Liu, Linfeng Zhang
arXiv Machine Learning
Jul 2

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

arXiv:2601. 11641v3 Announce Type: replace-cross Abstract: While Diffusion Transformers (DiTs) have achieved notable progress in video generation, this long-sequence generation task remains constrained by the quadratic complexity inherent to self-attention mechanisms, creating significant barriers to practical deployment.

By Yuxi Liu, Yipeng Hu, Zekun Zhang, Kunze Jiang, Kun Yuan
arXiv Computer Vision
Oct 16

Sparse Attention for Video Generation Acceleration via Growing Sparsity and Reduced Search

The paper introduces Growing Sparsity and Reduced Search (GSRS), a training‑free framework that accelerates Diffusion Transformers for video generation by progressively increasing sparsity and reusing early attention blocks to limit search space. GSRS uses Top‑p selection for adaptive sparsity and a precise evaluation scheme that computes exact attention weights with negligible overhead. Experiments on HunyuanVideo and Wan 2.1 show 1.55×–2.21× speedups while preserving generation quality comparable to full attention.

By Yun Dai, Yuyun Chen, Jiarui Wen, Huiping Zhuang, Cen Chen, Ziqian Zeng
arXiv Computer Vision
Oct 5

Moving Forward with Video Saliency: A New Dataset and Benchmark where Motion Matters

The paper introduces SalTempto, a new video saliency benchmark featuring 224 highly dynamic clips from HACS‑Segments, each with event context and gaze data from up to 16 subjects. It shows that static baselines perform poorly on SalTempto compared to LEDOV, while fine‑tuned temporal models achieve significant gains, yet still leave a large portion of explainable gaze information unaccounted for. The dataset highlights the need for better temporal modeling in video saliency research.

By Susmit Agrawal, Rebecca Wanner, Juliane Verwiebe, Matthias Tangemann, Matthias Bethge, Matthias K\"ummerer
arXiv AI
3d ago

VMonarch: Efficient Video Diffusion Transformers with Structured Attention

VMonarch introduces a structured Monarch matrix to replace the quadratic attention in Video Diffusion Transformers, enabling sub‑quadratic computation. The method captures intra‑ and inter‑frame correlations, mitigates instability artifacts, and integrates an online entropy algorithm with FlashAttention for efficient updates. Experiments show comparable or better generation quality, a 17.5× reduction in FLOPs, and over 5× speedup for long videos, outperforming state‑of‑the‑art sparse attention at 90% sparsity.

By Cheng Liang, Haoxian Chen, Liang Hou, Qi Fan, Gangshan Wu, Xin Tao, Limin Wang