The authors reformulate caching-based diffusion acceleration as revisable trajectory inference.
Revising skipped diffusion steps with later feedback improves caching efficiency and quality.
PoSTCache treats cached predictions as provisional, correcting them with signals from later full-computation steps and selecting anchors by expected uncertainty reduction.
Top University
Liwei Lan · Hui Chen · Jungong Han · Guiguang Ding
Tsinghua University
Research Digest··3 min read
The authors propose PoSTCache, a training-free framework that makes diffusion caching postdictive: predictions for skipped steps are first forecast, then revised using discrepancies observed at later full-computation steps.
§