CtrlCache cuts interactive video world model compute up to 1.41x

A new Hugging Face paper argues that interactive video world models leave an obvious scheduling signal on the table: the user's next controls arrive before the next chunk is denoised, so the system already knows when the view is about to swing. CtrlCache turns that into a training-free caching policy and reports 1.21x to 1.41x DiT-backbone speedups on three backbones with no retraining.

The mechanism is a four-state label on each video chunk. The abstract spells it out: the scheduler "detects action changes across and within chunks, and labels each chunk as initial, transition, turning, or steady state," and at one interior denoising step "initial and transition chunks retain full computation, while turning and steady chunks reuse the transformer residual from the most recent fully computed step in the same chunk." A separate frequency-mixed history prior reuses low-frequency structure from the preceding clean latent during steady chunks, where the scene layout barely moves.

On the WBench navigation track, Matrix-Game 2.0 goes from an Original Overall of 0.6543 to 0.6786 under full CtrlCache at a 1.41x speedup, with LingBot-World v1 and v2 landing at 1.26x and 1.21x. The authors claim CtrlCache is "the only method that improves over Original on all five WBench dimensions" on Matrix-Game 2.0, and that caching baselines remain marginally faster because "their reuse criteria are free to skip computation at action transitions" where CtrlCache instead spends it. The paper joins a steady run of video-generation research we have tracked — 44 stories in the last 90 days — and like most of them reports latency and quality in paired numbers rather than end-to-end wall-clock wins.