SourceMask: Provenance-Aware Solver Control for Diffusion Caching

Diffusion caching replaces expensive network evaluations with reused or predicted values. Existing cache policies determine when to refresh and what to return, while the numerical sampler usually consumes fresh and approximate evaluations in the same way. We introduce SourceMask, a provenance-aware interface for UniPC: the current corrector is enabled when both classifier-free guidance branches are fresh, and every returned evaluation is retained in predictor history. This rule adds no denoiser call. A complete-step analysis identifies how current-corrector and predictor-history paths jointly transmit cache error. On 944 Wan video prompts, SourceMask reduces LPIPS to the matched full-compute trajectory by 6.16%, with improvements on 935 prompts at the same evaluation budget. Separate studies establish gains for direct reuse, output prediction, and held-out SeaCache/SenCache combinations. A predictor-matched extension further controls UniP's high-order residual using nested-history agreement. Experiments across solver orders, sampling grids, and a second video model characterize where these controls help. The results establish evaluation provenance as a practical signal for coordinating diffusion caches with predictor–corrector solvers.