The 0.1-second trap: ComfyUI caches identical jobs

Failed, then fixed2026-10-23 · SDXL 1.0 · all

Bar chart comparing 0.1 seconds for a repeated seed with 8.8 seconds when the seed changes — The same GPU and model gave 0.1 s when the seed was repeated (a cache hit) and 8.8 s when each run used a new seed.

Our first warm runs reported about 0.1 s per image. That was not a fast GPU.

Why

ComfyUI caches results. When the same workflow is sent again with the same seed, it returns the cached image instead of generating a new one.

Fix

Every warm run now uses a different seed (seed + run number). The cold run is measured separately. The first, invalid result file is kept, clearly marked, so the mistake stays visible.