Here's a thread on how KV Caching is incorporated in QwenImage 2.1 👇
It separates the fixed context from image positions that change at every denoising step. Cache the fixed part once, and later steps get much cheaper.
We measured a 2.55x speedup 💥
1/8