AI RESEARCH

SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing

arXiv CS.CV

ArXi:2603.08982v1 Announce Type: new Diffusion Transformers (DiTs) have become a leading backbone for video generation, yet their quadratic attention cost remains a major bottleneck. Sparse attention reduces this cost by computing only a subset of attention blocks. However, prior methods often either drop the remaining blocks, which incurs information loss, or rely on learned predictors to approximate them,