Apple researchers say text-to-video models can run faster by avoiding attention calculations that rarely affect the final output. Their CalibAtt method identifies stable sparsity patterns in a model’s spatiotemporal attention and compiles them into optimized operations before inference.
Modern video generation systems spend substantial compute on attention, the mechanism that decides how tokens relate across space and time. Apple’s paper reports that many token-to-token connections consistently receive negligible scores, and that some patterns repeat across queries and local token blocks.
CalibAtt performs an offline calibration pass, then skips selected connections during generation while computing the remaining ones densely. The researchers tested it on Wan 2.1 14B, Mochi 1, and several distilled models at different resolutions. They report up to 1.58x end-to-end speedup while preserving video quality and text-video alignment.
The method is training-free, which makes it more practical than approaches that require retraining a model. Its usefulness will still depend on hardware support and whether the same sparse patterns hold across broader production workloads.