Sayak Paul: KV caching in QwenImage 2.1 measured at 2.55x speedup

Sayak Paul (@RisingSayak) posted a thread on how KV caching is incorporated in QwenImage 2.1. The method separates the fixed context from image positions that change at every denoising step, caches the fixed part once, and makes later steps cheaper.

He said they measured a 2.55x speedup.