Research map / Video / Video generation
AdaCluster: Adaptive Query-Key Clustering for Sparse Attention in Video Generation
Haoyue Tan, Shengnan Wang, Yulin Qiao, Juncheng Zhang, Youhui Bai, Ping Gong, Zewen Jin, Cheng Li
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add per-layer adaptive cluster counts instead of equal counts (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | PSNR: 29.007 -> 30.580 | Improved |
| add per-layer adaptive cluster counts instead of equal counts (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | SSIM: 0.724 -> 0.835 | Improved |
| add per-layer adaptive cluster counts instead of equal counts (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | LPIPS: 0.378 -> 0.203 | Improved |
| add TensorQuest cluster selection instead of mean-based selection (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | PSNR: 28.941 -> 30.580 | Improved |
| add TensorQuest cluster selection instead of mean-based selection (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | SSIM: 0.687 -> 0.835 | Improved |
| add TensorQuest cluster selection instead of mean-based selection (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | LPIPS: 0.410 -> 0.203 | Improved |
| add query normalization (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | PSNR: 29.56 -> 30.58 | Improved |
| add query normalization (ablation) | AdaCluster on HunyuanVideo · HunyuanVideo 1280x720 | LPIPS: 0.317 -> 0.203 | Improved |
| raise sparsity and initial TopK (83.2%, TopK 32 setting) | AdaCluster on HunyuanVideo · HunyuanVideo hyperparameter grid | PSNR / speedup: 30.58 -> 28.82 PSNR, 1.68x -> 1.90x | Mixed |
| lower KV threshold to 3.5 (77.5% sparsity) | AdaCluster on HunyuanVideo · HunyuanVideo hyperparameter grid | PSNR / speedup: 30.58 -> 31.88 PSNR, 1.68x -> 1.69x | Mixed |
| use AdaCluster instead of SVG2 | AdaCluster · Wan2.1-14B on H100 | speedup: 1.61x -> 1.81x | Improved |
| use AdaCluster instead of SpargeAttn and SVG2 at 28.3K tokens | AdaCluster · HunyuanVideo efficiency | speedup: 1.28x, 1.46x -> 1.53x | Improved |
Closest papers by meaning
- HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion (NeurIPS 2026 · similarity 0.92)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.91)
- iCATS: Fast Video Generation via Interaction-Aware Sparse Attention and Timestep-Adaptive Sparsity (NeurIPS 2026 · similarity 0.89)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.88)
- Sparse VideoGen2: Accelerate Video Generation with Sparse Attention via Semantic-Aware Permutation (NeurIPS 2025 · similarity 0.88)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.87)
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.
