SyncCache speeds up AI portrait animation up to 4.12x with no quality loss

SyncCache tackles the high computational cost of Diffusion Transformers (DiTs) in audio-driven portrait animation — models that convert speech into a talking face video. Existing training-free caching methods were designed for text-conditioned generation and fail to account for the spatial and modality imbalances inherent in portrait animation: the audio-driven mouth and facial movements change rapidly, while the background remains static.

SyncCache introduces two innovations: Spatially-Asymmetric Probing, which allocates more cache granularity to dynamic human regions, and Modality-Decoupled Caching, which bypasses heavy DiT blocks by reusing stable inter-block residuals while continuously recomputing lightweight audio blocks to keep lip sync accurate. A cache ratio is tuned via offline dynamic programming, and the method requires no fine-tuning. Experiments on HunyuanVideo-Avatar and Wan-S2V show up to 4.12x and 3.75x speedups respectively, with no measurable drop in visual quality or audio alignment. The paper was accepted at ECCV 2026.