CtrlCache: Accelerating Interactive Video World Models with Control-Aware Caching
CtrlCache uses the user’s control sequence to decide when video world models can safely reuse computation.
The paper says action changes are where structural similarity drops, while low-frequency structure tends to persist. CtrlCache labels chunks by control state, keeps full computation for initial and transition chunks, and reuses transformer residuals for turning and steady chunks at a selected denoising step. It also adds frequency-mixed history prior guidance without another DiT forward pass. In tests on Matrix-Game 2.0 and LingBot-World v1/v2, the authors report 1.21x to 1.41x DiT-backbone speedups with no retraining and higher WBench Overall scores. HF Daily Papers' note
The paper says action changes are where structural similarity drops, while low-frequency structure tends to persist. CtrlCache labels chunks by control state, keeps full computation for initial and transition chunks, and reuses transformer residuals for turning and steady chunks at a selected denoising step. It also adds frequency-mixed history prior guidance without another DiT forward pass. In tests on Matrix-Game 2.0 and LingBot-World v1/v2, the authors report 1.21x to 1.41x DiT-backbone speedups with no retraining and higher WBench Overall scores. HF Daily Papers' note
score 4