EchoAttention: Exploiting Token-Pair Redundancy and Frame-Block Similarity for Efficient Video Generation

Diffusion Transformers (DiTs) are increasingly adopted for long-video generation, yet inference is dominated by the quadratic cost of 3D full attention. Sparse attention mitigates this bottleneck by exploiting *token-pair redundancy* and pruning query-key interactions. Nevertheless, its effectiveness on video generation is often constrained by non-sparse attention heads, making it hard to strike a good balance between inference speed and generation quality. To address this, we identify another pervasive but overlooked redundancy specific to video DiTs: *frame-block similarity*, where frame-blocks in attention weights exhibit highly similar distributions and can be well approximated by lightweight linear calibration. Motivated by this observation, we propose **EchoAttention**, which jointly leverages *token-pair redundancy* (*Sparse* operator) and *frame-block similarity* (*Echo* operator), together with a fine-grained routing policy learned via three-stage distillation. This design enables efficient handling of both sparse and non-sparse heads, overcoming the inherent ceiling of purely sparse attention and yielding a better speed-quality trade-off. Across public video DiTs, EchoAttention consistently improves the speed-quality frontier over SOTA sparse-attention baselines, reducing end-to-end latency up to 2.42$\times$ with minimal quality loss.

Video generation models can create high-quality videos, but they are slow and computationally expensive because they repeatedly compare many parts of a video with one another. This cost becomes especially large for longer or higher-resolution videos. This paper introduces EchoAttention, a method that speeds up video generation by reducing repeated computation. We observe that many video frames contain similar patterns, so the model can reuse representative information from some frames and make small corrections instead of recomputing everything. EchoAttention also automatically decides where this reuse is appropriate. Experiments on public video generation models show that EchoAttention significantly reduces generation time while keeping video quality close to the original models. This can make video generation faster, cheaper, and more energy-efficient.