Research map / Video / Video generation
RAPID: Reusing Attention Sparsity with Inter-step Adaptation for Efficient Video Diffusion
Shangran Lin, Lu Lu, Jian Chen, Qiang Liu
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 Top-K anchor to Top-p block selection (hybrid vs Top-p only) (ablation) | RAPID · Wan2.1-14B, density 39.2% | PSNR: 18.922 -> 25.897 | Improved |
| add Top-p expansion to Top-K block selection (hybrid vs Top-K only) (ablation) | RAPID · Wan2.1-14B, density 39.2% | PSNR: 25.487 -> 25.897 | Improved |
| add Top-p expansion to Top-K block selection (hybrid vs Top-K only) (ablation) | RAPID · Wan2.1-14B, density 54.4% | PSNR: 27.377 -> 28.176 | Improved |
| extend dense warm-up from 25% to 40% of steps (ablation) | RAPID · Wan2.1-14B | PSNR/quality | No clear effect |
| lengthen dense warm-up from 5% to 25% of steps (ablation) | RAPID · Wan2.1-14B | generation quality | Improved |
| raise Top-p threshold tau from 0.5 to 0.9 (k_min fixed at 10%) (ablation) | RAPID · Wan2.1-14B, Tw=25% | generation quality | Improved |
| reuse one-shot cached mask instead of per-step dynamic masks, at matched density (RAPID vs Radial Attn) | Wan2.1-14B · VBench-2.0 prompts (768P, 81 frames) | PSNR: 22.916 -> 26.112 | Improved |
| reuse one-shot cached mask on HunyuanVideo (RAPID vs Radial Attn) | HunyuanVideo · VBench-2.0 prompts (768P, 129 frames) | PSNR: 27.499 -> 31.488 | Improved |
| add Turbo mode: re-threshold cached scores for more aggressive mask in later steps | RAPID · Wan2.1-14B | speedup / PSNR: 1.53x -> 1.79x speedup; 26.112 -> 22.612 PSNR | Mixed |
| add Turbo mode to RAPID on HunyuanVideo | RAPID · HunyuanVideo | speedup / PSNR: 1.73x -> 2.01x speedup; 31.488 -> 26.786 PSNR | Mixed |
Closest papers by meaning
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.93)
- Re-ttention: Ultra Sparse Visual Generation via Attention Statistical Reshape (NeurIPS 2025 · similarity 0.92)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.92)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.92)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.91)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.91)
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.
