Flash-BoN: cheap drafts beat guided search for diffusion inference-time scaling, +8% AUC at scale

gowthami_s · x · 2026-09-10

Gowthami Somepalli's team presents Flash-BoN at ECCV, showing that wall-clock cost—not NFE count—is the right yardstick for diffusion inference-time scaling.

Under matched end-to-end runtime, simple Best-of-N already matches or beats several guided search methods, since repeated intermediate verification eats the time budget that could explore new candidates.
Flash-BoN jointly optimizes three acceleration knobs—timestep truncation, layer skipping and activation proxies—once per model to mass-produce cheap drafts, then prunes with pointwise scores, head-to-head compares survivors, and refines the winner at full quality.
Across 9 model-benchmark combinations, Flash-BoN leads normalized AUC everywhere, with gains growing to +8% AUC at larger model scales.
The idea transfers to RL post-training: draft 16 rollouts, train on the top 6 and bottom 2.

Original post →

Flash-BoN: cheap drafts beat guided search for diffusion inference-time scaling, +8% AUC at scale

More from Multimodal