CtrlCache: Accelerating Interactive Video World Models with Control-Aware Caching
CtrlCache speeds interactive video world models 1.21x to 1.41x with training-free, control-aware caching.
CtrlCache is a training-free caching framework for interactive video world models that generate chunks autoregressively with few-step denoising. Because controls arrive before denoising, it detects action changes and labels each chunk as initial, transition, turning, or steady, keeping full computation for initial and transition chunks while reusing a transformer residual for turning and steady chunks. A frequency-mixed history prior reuses persistent low-frequency structure from the previous clean latent without another DiT forward pass. On Matrix-Game 2.0 and LingBot-World v1/v2 it reports 1.21x to 1.41x DiT-backbone speedups and higher WBench Overall scores than original inference.
- Training-free caching schedules compute from action changes before denoising.
- Chunks are labeled initial, transition, turning, or steady state.
- Turning and steady chunks reuse a transformer residual from a full step.
- Reports 1.21x to 1.41x speedups and higher WBench scores without retraining.
- AI models
- Matrix-Game 2.0LingBot-World
From the original article
1 min read at huggingface.co
Interactive video world models need to generate each video chunk efficiently while responding faithfully to user controls. Many systems use chunk-wise autoregressive generation with few-step denoising, but each chunk still requires several costly denoising iterations.
By Shangye Song, Dong Gong, Hong Jia · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.