Research map / Video / Video generation
EchoAttention: Exploiting Token-Pair Redundancy and Frame-Block Similarity for Efficient Video Generation
Yifei Xia, Fangcheng Fu, Hao Yuan, Suhan Ling, Xupeng Miao, Huixia Li, Yuxi Ren, Xin Xia, Xuefeng Xiao, Bin CUI
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 Echo operator alongside Sparse operator (vs sparse-only) (ablation) | EchoAttention · Wan2.1-1.3B | VBench / PSNR: VBench 81.5 -> 83.2; PSNR 17.02 -> 19.38 | Improved |
| add Sparse operator alongside Echo operator (vs echo-only) (ablation) | EchoAttention · Wan2.1-1.3B | VBench / PSNR: VBench 81.3 -> 83.2; PSNR 16.72 -> 19.38 | Improved |
| use Echo operator instead of full attention for routed heads (ablation) | EchoAttention · Wan2.1-1.3B | latency speedup: 1.49x -> 1.97x, VBench 83.0 -> 83.2 | Improved |
| learned routing instead of manual profiling-based routing (ablation) | EchoAttention · Wan2.1-1.3B | PSNR: 18.33 -> 19.38 | Improved |
| three-stage distillation instead of always soft routing (ablation) | EchoAttention · Wan2.1-1.3B | VBench: 82.5 -> 83.2 | Improved |
| three-stage distillation instead of always straight-through (ablation) | EchoAttention · Wan2.1-1.3B | VBench: 82.6 -> 83.2 | Improved |
| add hard-routing stage to schedule (ablation) | EchoAttention · Wan2.1-1.3B | VBench: 82.8 -> 83.2 | Improved |
| vary soft and st schedule hyperparameters (ablation) | EchoAttention · Wan2.1-1.3B | VBench: 83.1-83.2 | No clear effect |
| fixed 0-index prototype instead of random prototype (ablation) | EchoAttention · Wan2.1-1.3B | VBench: 82.9 -> 83.2 | Improved |
| adaptive top-1 similarity prototype instead of fixed 0-index (ablation) | EchoAttention · Wan2.1-1.3B | VBench: 83.2 -> 83.0 | No clear effect |
| EchoAttention vs best static sparse baseline SLA | Wan2.1-1.3B · Wan2.1-1.3B | speedup: 1.81x -> 1.97x | Improved |
Closest papers by meaning
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.90)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.90)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.89)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.89)
- EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs (NeurIPS 2026 · similarity 0.89)
- FPSAttention: Training-Aware FP8 and Sparsity Co-Design for Fast Video Diffusion (NeurIPS 2025 · similarity 0.89)
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.
