Definition
Plain language
A trick that skips redundant work in image generation by reusing a cached intermediate result on most diffusion steps.
As stated in the literature
In VibeServe's Show-o2 stack, an optimization that skips the unconditional CFG branch on most diffusion steps and reuses a cached unconditional vector to cut compute.
Why it matters: It roughly halves the compute cost of guidance with minimal quality loss, which matters when serving image generation at scale.
For example, instead of computing the unconditional branch at every diffusion step, the system reuses the cached one for several steps in a row before refreshing it.
Heard on the show
“The other big win — they call it the CFG stride trick — was realizing they could skip the unconditional branch on most diffusion steps and reuse a cached vector instead.”Episode 027 — When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure