AI Paper Atlas

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

ICML 2026 · Video › Video generation · approach: Efficient video diffusion models

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
add Echo operator alongside Sparse operator (vs sparse-only) (ablation)EchoAttention · Wan2.1-1.3BVBench / PSNR: VBench 81.5 -> 83.2; PSNR 17.02 -> 19.38Improved
add Sparse operator alongside Echo operator (vs echo-only) (ablation)EchoAttention · Wan2.1-1.3BVBench / PSNR: VBench 81.3 -> 83.2; PSNR 16.72 -> 19.38Improved
use Echo operator instead of full attention for routed heads (ablation)EchoAttention · Wan2.1-1.3Blatency speedup: 1.49x -> 1.97x, VBench 83.0 -> 83.2Improved
learned routing instead of manual profiling-based routing (ablation)EchoAttention · Wan2.1-1.3BPSNR: 18.33 -> 19.38Improved
three-stage distillation instead of always soft routing (ablation)EchoAttention · Wan2.1-1.3BVBench: 82.5 -> 83.2Improved
three-stage distillation instead of always straight-through (ablation)EchoAttention · Wan2.1-1.3BVBench: 82.6 -> 83.2Improved
add hard-routing stage to schedule (ablation)EchoAttention · Wan2.1-1.3BVBench: 82.8 -> 83.2Improved
vary soft and st schedule hyperparameters (ablation)EchoAttention · Wan2.1-1.3BVBench: 83.1-83.2No clear effect
fixed 0-index prototype instead of random prototype (ablation)EchoAttention · Wan2.1-1.3BVBench: 82.9 -> 83.2Improved
adaptive top-1 similarity prototype instead of fixed 0-index (ablation)EchoAttention · Wan2.1-1.3BVBench: 83.2 -> 83.0No clear effect
EchoAttention vs best static sparse baseline SLAWan2.1-1.3B · Wan2.1-1.3Bspeedup: 1.81x -> 1.97xImproved

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