AIQwenImage 2.1 reportedly gets a 2.55x speedup from KV caching
A user says the approach caches fixed context once, keeping it separate from image positions that change at every denoising step.
TLDR
A user reports measuring a 2.55x speedup with KV caching in QwenImage 2.1. Their thread describes separating fixed context from image positions that change during denoising. Caching the fixed part once makes later steps cheaper, the author says.
QwenImage 2.1 reportedly gets a 2.55x speedup from KV caching
A user says the approach caches fixed context once, keeping it separate from image positions that change at every denoising step.
