A ComfyUI node named TE-Speed-QwenImage21 accelerates QwenImage2.1 inference. It uses output prediction to reduce processing time.
The node offers two modes: standard (step_cache=te_predictor) for one continuous step, and speed (step_cache=speed) for two steps, which may increase quality deviation risk. The default reuse_threshold is 0.06.
It should be placed in the workflow after UNETLoader and QwenImage21Cache (optional), then before KSampler. The node does not replace the official Prefix KV Cache and aims to reduce time without changing the sampling schedule.
Highlights
- 默认 reusethreshold=0.06，可在节点中调节速度与画质平衡,越大越快,目前已为测试最佳值。
- stepcache=tepredictor 为标准模式，最多连续预测 1 步。
- stepcache=speed 为速度模式，最多连续预测 2 步；连续预测会增加画质偏差风险。
- speed 会在标准模式上增加一次连续输出预测，实际加速取决于采样器、分辨率和阈值设置。